PROCEEDINGS SIMBig 2015 - Organizing Committee SIMBig 2015 - Program Committee SIMBig 2015 - List of Contributions SIMBig 2015 - Keynote Speaker Presentations SIMBig 2015 - Sessions, Papers The SIMBig 2015 Organizing Committee con rms that full and concise papers accepted for this publication: • Meet the de nition of research in relation to creativity, originality, and increasing humanity's stock of knowledge; • Are selected on the basis of a peer review process that is independent, quali ed expert review; • Are published and presented at a conference having national and international signi cance as evidenced by registrations and participation; and • Are made available widely through the Conference web site.
Disclaimer: The SIMBig 2015 Organizing Committee accepts no responsibility for omissions and errors. Organizing Committee of SIMBig 2015 General Organizers • Juan Antonio
LIRMM, FRANCE
LOSSIO-VENTURA, University of Montpellier • Hugo ALATRISTA-SALAS, Ponti cia Universidad Catolica del Peru
GRPIAA Labs, PERU
• Cristhian GANVINI VALCARCEL Universidad Andina del Cusco, PERU • Armando FERMIN PEREZ, Universidad Nacional Mayor de San Marcos,
PERU
WTI (Web and Text Intelligence Track) • Jorge Carlos VALVERDE-REBAZA, University of Sa~o Paulo, Brazil • Alneu DE ANDRADE LOPES, ICMC { University of S~ao Paulo, Brazil • Ricardo Bastos CAVALCANTE PRUDENCIO, Federal University of Pernambuco, Brazil • Estevam HRUSCHKA, Federal University of Sa~o Carlos, Brazil • Ricardo CAMPOS, Polytechnic Institute of Tomar - LIAAD{INESC Technology and Science, Portugal SIMBig 2015 Program Committee • Nathalie Abadie, French National Mapping Agency, COGIT, FRANCE • Elie Abi-Lahoud, University College Cork, Cork, IRELAND • Salah AitMokhtar, Xerox Research Centre Europa, FRANCE • Sophia Ananiadou, NaCTeM - University of Manchester, UNITED KING
DOM • Marcelo Arenas, Ponti cia Universidad Catolica de Chile, CHILE • Jero^me Aze, LIRMM - University of Montpellier, FRANCE • Pablo Barcelo, Universidad de Chile, CHILE • Cesar A. Beltran Castan~on, GRPIAA - Ponti cal Catholic University of
Peru, PERU • Albert Bifet, Huawei Noah's Ark Research Lab, Hong Kong, CHINA • Sandra Bringay, LIRMM - Paul Valery University, FRANCE • Oscar Corcho, Ontology Engineering Group - Polytechnic University of
Madrid, SPAIN • Gabriela Csurka, Xerox Research Centre Europa, FRANCE • Frederic Flouvat, PPME Lab - University of New Caledonia, NEW CALE
DONIA • Andre Freitas, Dept. of Computer Science and Mathematics, University of Passau, GERMANY • Adrien Guille, ERIC Lab - University of Lyon 2, FRANCE • Hakim Hacid, Zayed University, UNITED ARAB EMIRATES • Sebastien Harispe, LGI2P/EMA Research Centre, Site EERIE, Parc Scienti que, FRANCE • Dino Ienco, Irstea, FRANCE • Diana Inkpen, University of Ottawa, CANADA • Clement Jonquet, LIRMM - University of Montpellier, FRANCE • Al pio Jorge, Universidade do Porto, PORTUGAL • Yannis Korkontzelos, NaCTeM - University of Manchester, UNITED KINGDOM • Eric Kergosien, GERiCO Lab - University of Lille 3, FRANCE • Peter Mika, Yahoo! Research Labs - Barcelone, SPAIN • Phan Nhat Hai, Oregon State University, UNITED STATES of AMERICA • Jordi Nin, Barcelona Supercomputing Center (BSC) - BarcelonaTECH,
SPAIN • Miguel Nun~ez del Prado Cortez, Intersec Lab - Paris, FRANCE • Thomas Opitz, Biostatistics and Spatial Processes - INRA, FRANCE • Yoann Pitarch, IRIT - Toulouse, FRANCE • Pascal Poncelet, LIRMM - University of Montpellier, FRANCE • Julien Rabatel, LIRMM, FRANCE • Jose Luis Redondo Garc a, EURECOM, FRANCE • Mathieu Roche, Cirad - TETIS - LIRMM, FRANCE • Nancy Rodriguez, LIRMM - University of Montpellier, FRANCE • Arnaud Sallaberry, LIRMM - Paul Valery University, FRANCE • Nazha Selmaoui-Folcher, PPME Labs - University of New Caledonia,
NEW CALEDONIA • Maguelonne Teisseire, Irstea - LIRMM, FRANCE • Paulo Teles, LIAAD-INESC Porto LA, Porto University, PORTUGAL • Julien Velcin, ERIC Lab - University of Lyon 2, FRANCE • Maria-Esther Vidal, Universidad Simon Bol var, VENEZUELA • Boris Villazon-Terrazas, Expert System Iberia - Madrid, SPAIN • Osmar R. Zaane, Department of Computing Science, University of Alberta, CANADA WTI 2015 Program Committee (Web and Text Intelligence Track) • Adam Jatowt, Kyoto University, Japan, • Claudia Orellana Rodriguez, Insight Centre for Data Analytics - University College Dublin, Ireland, • Ernesto Diaz-Aviles, IBM Research, Ireland, • Jannik Strotgen, Heidelberg University, Germany, • Jo~ao Paulo Cordeiro, University of Beira Interior, Portugal, • Lucas Drumond, University of Hildesheim, Germany, • Leon Derczynski, University of She eld, UK, • Miguel Martinez-Alvarez, Signal, UK, • Brett Drury, USP, Brazil, • Celso Anto^nio Alves Kaestner, UTFPR, Brazil, • Edson Matsubara, UFMS, Brazil, • Flavia Barros, UFPe, Brazil, • Hercules Antonio do Prado, PUC Bras lia, Brazil, • Huei Lee, UNIOESTE, Brazil, • Jo~ao Lu s Rosa, USP, Brazil, • Jose Lu s Borges, Universidade do Porto, Portugal, • Jose Paulo Leal, Universidade do Porto, Portugal, • Maria Cristina Ferreira de Oliveira, USP, Brazil, • Ronaldo Prati, UFABC, Brazil, • Thiago A. S. Pardo, USP, Brazil, • Solange Rezende, USP, Brazil, • Marcos Aurelio Domingues, USP, Brazil. List of contributions 1 2
Overview of SIMBig 2015 10 Overview of SIMBig 2015: 2nd Annual International Symposium on Information Management and Big Data spaJuan Antonio Lossio-Ventura and Hugo Alatrista-Salas . . . . . . . 10
Overview of SIMBig 2015: 2nd Annual International Symposium on Information Management and Big Data Juan Antonio Lossio-Ventura Hugo Alatrista-Salas LIRMM, University of Montpellier Pontificia Universidad Cato´lica del Peru´
Montpellier, France Lima, Peru juan.lossio@lirmm.fr halatrista@pucp.pe Big Data is a popular term used to describe the exponential growth and availability of both structured and unstructured data. The aim of the symposium is to present the analysis methods for managing large volumes of data through techniques of artificial intelligence and data mining.
Bringing together main national and international actors in the decision-making field to state in new technologies dedicated to handle large amount of information. 1 Big Data is a popular term used to describe the exponential growth and availability of both structured and unstructured data. This has taken place over the last 20 years. For instance, social networks such as Facebook, Twitter and Linkedin generate masses of data, which is available to be accessed by other applications. Several domains, including biomedicine, life sciences and scientific research, have been affected by Big Data1. Therefore there is a need to understand and exploit this data.
This process can be carried out thanks to “Big Data Analytics” methodologies, which are based on Data Mining, Natural Language Processing, etc. That allows us to gain new insight through data-driven research (Madden, 2012; Embley and Liddle, 2013). A major problem hampering Big Data Analytics development is the need to process several types of data, such as structured, numeric and unstructured data (e.g. video, audio, text, image, etc)2.
Therefore, the second edition of the Annual International Symposium on Information Management and Big Data - SIMBig 20153, aims to present the analysis methods for managing large volumes of data through techniques of artificial intelligence and data mining. Counting with main national and
1By 2015 the average of data annually generated in hospitals is 665TB: http://ihealthtran.com/wordpress/2013/ 03/infographic-friday-the-body-as-a-source-of-big-data/.
2Today, 80% of data is unstructured such as images, video, and notes 3http://simbig.org/SIMBig2015/ international actors in the decision-making field to state in new technologies dedicated to handle large amount of information.
Our first edition, SIMBig 20144 took place in Cuzco Peru too in September 2015. SIMBig 2014 has been indexed on DBLP5 (Lossio-Ventura and Alatrista-Salas, 2014) and on CEUR Workshop Proceedings6. 1.1
SIMBig 2015 second edition has welcomed five keynote speakers experts in Big Data, Data Mining, Natural Language Processing (NLP), and Social Networks: • PhD. Pr. Albert Bifet, from HUAWEI
Noah’s Ark Lab, China; • PhD. Pr. Diana Inkpen, from University of
Ottawa, Canada; • PhD. Pr. Pascal Poncelet, from LIRMM Laboratory and University of Montpellier, France; • PhD. Pr. Mathieu Roche, from Cirad and
TETIS Laboratory, France; • PhD. Pr. Osmar R. Za¨ıane, from University of Alberta, Canada. 1.2
To share the new analysis methods for managing large volumes of data, we encouraged participation from researchers in all fields related to Big Data, Data Mining, and Natural Language Processing, but also Multilingual Text Processing, Biomedical NLP. Topics of interest of SIMBig 2015 included but were not limited to: • Big Data • Data Mining • Natural Language Processing 4https://www.lirmm.fr/simbig2014/ 5http://dblp2.uni-trier.de/db/conf/simbig/ simbig2014 6http://ceur-ws.org/Vol-1318/index.html • Bio NLP • Text Mining • Information Retrieval • Machine Learning • Semantic Web • Ontologies • Web Mining • Knowledge Representation and Linked Open
Data • Social Networks, Social Web, and Web Science • Information visualization • OLAP, Data Warehousing • Business Intelligence • Spatiotemporal Data • Health Care • Agent-based Systems • Reasoning and Logic • Constraints, Satisfiability, and Search 2
The academic goals of the symposium are varied, among which we can list the following: • Meet Latin American and foreign researchers, teachers, and students belonging to several domains of computer sciences, specially related to Big Data. • Promote the production of scientific articles, which will be evaluated by the international scientific community, in order to receive a feedback from experts. • Foster partnerships between Latin American universities, local universities and European universities. • Promote the creation of alliances between Peruvian universities, enabling decentralization of education. • Motivate students to learn more about computer sciences research to solve problems related to the management of information and Big Data. • Promote the research in Peruvian universities, mainly those belonging to the local organizing committee. • Create connections, forming networks of partnerships between companies and universities. • Promote the local and international tourism, in order to show to the participants the architecture, gastronomy and local cultural heritage. 3
Web and text intelligence are related areas that have been used to improve human computer interaction both in general and in particular to explore and analyze information that is available on the Internet. With the advent of social networks and the emergence of services such as Facebook, Twitter, and others, research in these areas has been greatly enhanced. In recent years, shared knowledge and experiences have established new and different types of personal and communal relationships which have been leveraged by social networks scientists to produce new insights. In addition there has been a huge increase in community activities on social networks.
The Web and Text Intelligence (WTI) track of SIMBig 2015 have provided a forum that brought together researchers and practitioners for exploring technologies, issues, experiences and applications that help us to understand the Web and to build automatic tools to better exploit this complex environment. The WTI track has fostered collaborations, exchange of ideas and experiences among people working in a variety of highly crossdisciplinary research fields such as computer science, linguistics, statistics, sociology, economics, and business.
The WTI track is a follow up of the 4th International Workshop on Web and Text Intelligence7, which took place in Curitiba, Brazil, October 2012, as a workshop of BRACIS 2012; the 3rd International Workshop on Web and Text Intelligence8, which took place in S˜ao Bernardo, Brazil, October 2010, as a workshop of SBIA10; the 2nd International Workshop on Web and Text Intelligence9, which took place in S˜ao Carlos, Brazil, September 2009, as a workshop of STIL09; the 1st Web and Network Intelligence10, which took place in Aveiro, Portugal, October 2009, as a thematic track of EPIA09; and the 1st International Workshop on Web and Text Intelligence, which took place in 7http://www.labic.icmc.usp.br/wti2012/ 8http://www.labic.icmc.usp.br/wti2010/ 9http://www.labic.icmc.usp.br/wti2009/ 10http://epia2009.web.ua.pt/wni/ Salvador, Brazil, October 2008, as a workshop of SBIA08. We want to thank our wonderful sponsors! We extend our sincere appreciation to our sponsors, without whom our symposium would not be possible. They showed their commitment to making our research communities more active. We invite you to support these community-minded organizations. 4.1
• Universit´e de Montpellier, France11 • Laboratoire de Informatique, Robotique et
Micro´electronique de Montpellier, France12 • Universidad Andina del Cusco, Peru´13 11http://www.umontpellier.fr/ 12http://www.lirmm.fr/ 13http://www.uandina.edu.pe/ 4.2
Collaborating Institutions • iMedia14 • Bioincuba15 • TechnoPark16 • Grupo de Reconocimiento de Patrones e In
teligencia Artificial Aplicada, PUCP, Peru´17 • Universidad Nacional Mayor de San Marcos,
Peru´18 • Escuela de Post-grado de la Pontificia Univer
sidad Cat´olica del Peru´19 14http://www.imedia.pe/ 15http://www.bioincuba.com/ 16http://technoparkidi.org/ 17http://inform.pucp.edu.pe/~grpiaa/ 18http://www.unmsm.edu.pe/ 19http://posgrado.pucp.edu.pe/la-escuela/ presentacion/ 20http://www.icmc.usp.br/Portal/ 21http://portal2.ipt.pt/ 22http://labic.icmc.usp.br/ 23http://www.inesctec.pt/liaad 24http://ppgcc.dc.ufscar.br/pesquisa/ laboratorios-e-grupos-de-pesquisa
25http://www2.ufscar.br/home/index.php Real-Time Big Data Stream Analytics Big Data is a new term used to identify datasets that we cannot manage with current methodologies or data mining software tools due to their large size and complexity. Big Data mining is the capability of extracting useful information from these large datasets or streams of data.
New mining techniques are necessary due to the volume, variability, and velocity, of such data. MOA is a software framework with classification, regression, and frequent pattern methods, and the new APACHE SAMOA is a distributed streaming software for mining data streams. 1 Introduction Big Data is a new term used to identify the datasets that due to their large size, we can not manage them with the typical data mining software tools.
Instead of defining “Big Data” as datasets of a concrete large size, for example in the order of magnitude of petabytes, the definition is related to the fact that the dataset is too big to be managed without using new algorithms or technologies. There is need for new algorithms, and new tools to deal with all of this data. Doug Laney (Laney, 2001) was the first to mention the 3 V’s of Big Data management: • Volume: there is more data than ever before, its size continues increasing, but not the percent of data that our tools can process • Variety: there are many different types of data, as text, sensor data, audio, video, graph, and more • Velocity: data is arriving continuously as streams of data, and we are interested in obtaining useful information from it in real time
Nowadays, there are two more V’s: • Variability: there are changes in the structure of the data and how users want to interpret that data • Value: business value that gives organizations a competitive advantage, due to the ability of making decisions based in answering questions that were previously considered beyond reach
For velocity, data stream real time analytics are needed to manage the data currently generated, at an ever increasing rate, from such applications as: sensor networks, measurements in network monitoring and traffic management, log records or click-streams in web exploring, manufacturing processes, call detail records, email, blogging, twitter posts and others. In fact, all data generated can be considered as streaming data or as a snapshot of streaming data, since it is obtained from an interval of time.
In the data stream model, data arrive at high speed, and algorithms that process them must do so under very strict constraints of space and time. Consequently, data streams pose several challenges for data mining algorithm design. First, algorithms must make use of limited resources (time and memory). Second, they must deal with data whose nature or distribution changes over time.
We need to deal with resources in an efficient and low-cost way. In data stream mining, we are interested in three main dimensions: • accuracy • amount of space necessary • the time required to learn from training ex
amples and to predict
These dimensions are typically interdependent: adjusting the time and space used by an algorithm can influence accuracy. By storing more precomputed information, such as look up tables, an algorithm can run faster at the expense of space.
An algorithm can also run faster by processing less information, either by stopping early or storing less, thus having less data to process. 2
MOA Massive Online Analysis (MOA) (Bifet et al., 2010) is a software environment for implementing algorithms and running experiments for online learning from evolving data streams. MOA includes a collection of offline and online methods as well as tools for evaluation. In particular, it implements boosting, bagging, and Hoeffding Trees, all with and without Na¨ıve Bayes classifiers at the leaves. Also it implements regression, and frequent pattern methods. MOA supports bidirectional interaction with WEKA, the Waikato Environment for Knowledge Analysis, and is released under the GNU GPL license. 3
APACHE SAMOA APACHE SAMOA (SCALABLE ADVANCED MASSIVE ONLINE ANALYSIS) is a platform for mining big data streams (Morales and Bifet, 2015). As most of the rest of the big data ecosystem, it is written in Java.
APACHE SAMOA is both a framework and a library. As a framework, it allows the algorithm developer to abstract from the underlying execution engine, and therefore reuse their code on different engines. It features a pluggable architecture that allows it to run on several distributed stream processing engines such as Storm, S4, and Samza.
This capability is achieved by designing a minimal API that captures the essence of modern DSPEs.
This API also allows to easily write new bindings to port APACHE SAMOA to new execution engines.
APACHE SAMOA takes care of hiding the differences of the underlying DSPEs in terms of API and deployment.
As a library, APACHE SAMOA contains implementations of state-of-the-art algorithms for distributed machine learning on streams. For classification, APACHE SAMOA provides a Vertical Hoeffding Tree (VHT), a distributed streaming version of a decision tree. For clustering, it includes an algorithm based on CluStream. For regression,
HAMR, a distributed implementation of Adaptive Model Rules. The library also includes metaalgorithms such as bagging and boosting.
The platform is intended to be useful for both research and real world deployments. 3.1
High Level Architecture We identify three types of APACHE SAMOA users: 1. Platform users, who use available ML algo
rithms without implementing new ones. 2. ML developers, who develop new ML algorithms on top of APACHE SAMOA and want to be isolated from changes in the underlying
SPEs. 3. Platform developers, who extend APACHE
SAMOA to integrate more DSPEs into
APACHE SAMOA. 4
Conclusions Big Data Mining is a challenging task, that needs new tools to perform the most common machine learning algorithms such as classification, clustering, and regression.
APACHE SAMOA is is a platform for mining big data streams, and it is already available and can be found online at http://www. samoa-project.net. The website includes a wiki, an API reference, and a developer’s manual.
Several examples of how the software can be used are also available.
Acknowledgments The presented work has been done in collaboration with Gianmarco De Francisci Morales, Bernhard Pfahringer, Geoff Holmes, Richard Kirkby, and all the contributors to MOA and APACHE SAMOA.
Detecting Locations from Twitter Messages
Invited Talk
Diana Inkpen
University of Ottawa
School of Electrical Engineering and Computer Science 800 King Edward Avenue, Ottawa, ON, K1N6N5, Canada
Diana.Inkpen@uottawa.ca 1
Extended Abstract There is a large amount of information that can be extracted automatically from social media messages. Of particular interest are the topics discussed by the users, the opinions and emotions expressed, and the events and the locations mentioned. This work focuses on machine learning methods for detecting locations from Twitter messages, because the extracted locations can be useful in business, marketing and defence applications (Farzindar and Inkpen, 2015).
There are two types of locations that we are interested in: location entities mentioned in the text of each message and the physical locations of the users. For the first type of locations (task 1), we detected expressions that denote locations and we classified them into names of cities, provinces/states, and countries. We approached the task in a novel way, consisting in two stages. In the first stage, we trained Conditional Random Field models with various sets of features. We collected and annotated our own dataset for training and testing. In the second stage, we resolved cases when more than one place with the same name exists, by applying a set of heuristics (Inkpen et al., 2015).
For the second type of locations (task 2), we put together all the tweets written by a user, in order to predict his/her physical location. Only a few users declare their locations in their Twitter profiles, but this is sufficient to automatically produce training and test data for our classifiers. We experimented with two existing datasets collected from users located in the U.S. We propose a deep learning architecture for the solving the task, because deep learning was shown to work well for other natural language processing tasks, and because standard classifiers were already tested for the user location task. We designed a model that predicts the U.S. region of the user and his/her U.S. state, and another model that predicts the longitude and latitude of the user's location. We found that stacked denoising autoencoders are well suited for this task, with results comparable to the state-of-the-art (Liu and Inkpen, 2015). 2
Biography Diana Inkpen is a Professor at the University of Ottawa, in the School of Electrical Engineering and Computer Science. Her research is in applications of Computational Linguistics and Text Mining. She organized seven international workshops and she was a program co-chair for the AI 2012 conference. She is in the program committees of many conferences and an associate editor of the Computational Intelligence and the Natural Language Engineering journals. She published a book on Natural Language Processing for Social Media (Morgan and Claypool Publishers, Synthesis Lectures on Human Language Technologies), 8 book chapters, more than 25 journal articles and more than 90 conference papers. She received many research grants, including intensive industrial collaborations.
Acknowledgments I want to thank my collaborators on the two tasks addressed in this work: Ji Rex Liu for his implementation of the proposed methods for both tasks and Atefeh Farzindar for her insight on the first task. I also thank the two annotators of the corpus for task 1, Farzaneh Kazemi and Ji Rex Liu.
This work is funded by the Natural Sciences and Engineering Research Council of Canada. Atefeh Farzindar and Diana Inkpen. 2015. Natural
Language Processing for Social Media. Morgan and Claypool Publishers. Synthesis Lectures on
Human Language Technologies.
Ji Liu and Diana Inkpen. 2015. Estimating User Location in Social Media with Stacked Denoising Autoencoders. In Proceedings of the 1st Workshop on Vector Space Modeling for Natural Language Processing, NAACL 2015, Denver, Colorado, pp. 201210. Opinion Mining: Taking into account the criteria!
Pascal Poncelet
LIRMM - UMR 5506 Campus St Priest - Baˆtiment 5
860 rue de Saint Priest 34095 Montpellier Cedex 5 - France
and UMR TETIS Montpellier - France
Pascal.Poncelet@lirmm.fr Today we are more and more provided with information expressing opinions about different topics. In the same way, the number of Web sites giving a global score, usually by counting the number of stars for instance, is also growing extensively and this kind of tools can be very useful for users interested by having a general idea. Nevertheless, sometimes the expressed score (e.g. the number of stars) does not really reflect what it is expressed in the text of a review. Actually, extracting opinions from texts is a problem that have been extensively addressed in the last decade and very efficient approaches are now proposed to extract the polarity of a text. In this presentation we focus on a topic related with opinion but rather than considering the full text we are interested with the opinions expressed on specific criteria. First we show how criteria can be automatically learnt. Second we illustrate how opinions are extracted.
By considering criteria we illustrate that it is possible to propose new recommender systems but also to evaluate how opinions expressed on the criteria evolve over time. 1 Extracting opinions that are expressed in a text is a topic that have been addressed extensively in the last decade (e.g. (Pang and Lee, 2008)). Usually proposed approaches mainly focus on the polarity of a text: this text is positive, negative or even neutral. Figure 1 shows an example of a review on a restaurant.
Actually this review has been scored quite well: 4 stars over 5. Any opinion mining tools will show that the review is much more negative than positive. Let us go deeper on this exemple. Even if We are here on a Saturday night and the food and service was amazing.
We brought a group back the next day and we were treated so poorly by a man with dark hair.
He ignored us when we needed a table for 6 to the point of us leaving to get takeaway.
Embarrassing and so disappointing. the review is negative it clearly illustrates that the reviewer was mainly disappointed by the service: he was in the Restaurant and found it amazing. We could imagine that, at that time, the service was not so bad. This exemple illustrates the problem we address in the presentation: we do not focus on a whole text rather we would like to extract opinions related to some specific criteria. Basically, by considering a set of user-specified criteria we would like to highlight (and obviously extract opinions) only on the relevant parts of the reviews focusing on these criteria. The paper is organized as follows. In Section 2 we give some ideas on how to automatically learn terms related to a criterium. We give also some clues for extracting opinions to the criteria in Section 3. Finally Section 4 concludes the paper. 2
Automatic extraction of terms related to a criterium First of all we assume that the end user is interested in a specific domain and some criteria.
Let us imagine that the domain is movie and the two criteria are actor and scenario. For each criterium we only need to have several keywords or terms of the criterium (seed of terms). For instance in the movie domain: Actor= {actor, acting, casting, character, interpretation, role, star} and Scenario={scenario, adaptation, narrative, original, scriptwriter, story, synopsis}. Intuitively two different sets may exist. The first one corresponding to all the terms that may be used for a criterium. Such a set is called a class. The second one corresponds to all the terms which are used in the domain but which are not in the class. This set is called anti-class. For instance the term theater is about movie but is not specific neither to the class actor nor scenario. Now the problem is to automatically learn the set of all terms for a class. Using experts or users to annotate documents is too expensive and error-prone. By the way there are many documents available on the internet having the terms of the criteria that can be learned. In a practical way by using a research engine it is easy and possible to get these documents.
For instance, the following query expressed in Google: ”+movie +actor -scenario adaptation narrative original -scriptwriter story -synopsis” will extract a set of documents of the domain movie (character +), having actor in the document and without (character −) scenario, adaptation, etc. In other terms we are able to automatically extract movie documents having terms only relative to the class actor. By performing some text preprocessing and taking into account a frequency of a term in a specific window of terms (see (Duthil et al., 2011) for a full description of the process as well as the measure that can be used to score the terms) we can extract quite relevant terms: the higher the score, the higher the probability of this term belonging to a class. Nevertheless as the number of documents to be analyzed is limited, some terms may not appear in the corpus. Usually these terms will have more or less the same score both in the class and the anti-class. They are called candidates and as we do not know the most appropriate class, a new query on the Web will extract new documents. Here again, a new score can be computed and all the terms with their associated scores can finally be stored in lexicons. Such lexicon can then be used to automatically segment a document for instance. 3
Extracting opinions A quite similar process may be adapted for extracted terms used to express opinions: adjectives, verbs and even grammatical patterns such as <adverb + adjective > in order to automatically learn positive and negative expressions. Then by using the new opinion lexicon extracted we can easily detect the polarity of a document. In the same way by using the segmentation performed in the previous step it is now possible to focus on criteria and then extract the opinion for a specific criterium. Interested reader may refer to (Duthil et al., 2012). 4 In the presentation we will present more in detail the main approach. Conducted experiments that will be presented during the talk will show that such an approach is very efficient when considering Precision and Recall measures. Furthermore some practical aspects will be addressed: how many documents? how many seed terms? the quality of the results for different domains? We will also show that such lexicons could also be very useful for recommending systems. For instance we are able to focus on the criteria that are addressed by newspapers and then recommend the end user only with a list of newspapers he/she could be interested in. In the same way, evaluating how opinions evolve over time on different criteria is of great interest for many different applications.
Interested reader may refer to (Duthil, 2012) for different applications that can be defined.
Acknowledgments The presented work has been done mainly during the Ph.D of Dr. Benjamin Duthil and in collaboration with Ge´rard Dray, Jacky Montmain, Michel Plantie´ from the Ecole des Mines d’Ale`s (France) and Mathieu Roche from the University of Montpellier (France).
References How to deal with heterogeneous data? The Big Data issue is traditionally characterized in terms of 3 V, i.e. volume, variety, and velocity.
This paper focuses on the variety criterion, which is a challenging issue.
Data heterogeneity and content In the context of Web 2.0, text content is often heterogenous (i.e. lexical heterogeneity).
For instance, some words may be shortened or lengthened with the use of specific graphics (e.g. emoticons) or hashtags. Specific processing is necessary in this context. For instance, with an opinion classification task based on the message SimBig is an aaaaattractive conference!, the results are generally improved by removing repeated characters (i.e. a).
But information on the sentiment intensity identified by the character elongation is lost with this normalization. This example highlights the difficulty of dealing with heterogeneous textual data content.
The following sub-section describes the heterogeneity according the document types (e.g. images and texts).
Heterogeneity and document types Impressive amounts of high spatial resolution satellite data are currently available. This raises the issue of fast and effective satellite image analysis as costly human involvement is still required.
Meanwhile, large amounts of textual data are available via the Web and many research communities are interested in the issue of knowledge extraction, including spatial information. In this context, image-text matching improves information retrieval and image annotation techniques (Forestier et al., 2012). This provides users with a more global data context that may be useful for experts involved in land-use planning (Alatrista Salas et al., 2014). 2
Text-mining method for matching heterogenous data A generic approach to address the heterogeneity issue consists of extracting relevant features in documents. In our work, we focus on 3 types of features: thematic, spatial, and temporal features. These are extracted in textual documents using natural language processing (NLP) techniques based on linguistic and statistic information (Manning and Schu¨tze, 1999): • The extraction of thematic information is based on the recognition of relevant terms in texts. For instance, terminology extraction techniques enable extraction of single-word terms (e.g. irrigation) or phrases (e.g. rice crops). The most efficient state-ofthe-art term recognition systems are based on both statistical and linguistic information (Lossio-Ventura et al., 2015). • Extracting spatial information from documents is still challenging. In our work, we use patterns to detect these specific named entities. Moreover, a hybrid method enables disambiguation of spatial entities and organizations. This method combines symbolic approaches (i.e. patterns) and machine learning techniques (Tahrat et al., 2013). • In order to extract temporal expressions in texts, we use rule-based systems like HeidelTime (Stro¨tgen and Gertz, 2010). This multilingual system extracts temporal expressions from documents and normalizes them.
HeidelTime applies different normalization strategies depending on the text types, e.g.
news, narrative, or scientific documents.
These different methods are partially used in the projects summarized in the following section.
More precisely, Section 3 presents two projects that investigate heterogeneous data in agricultural the domain.1 3
Applications in the agricultural domain New and emerging infectious diseases are an increasing threat to countries. Many of these diseases are related to globalization, travel and international trade. Disease outbreaks are conventionally reported through an organized multilevel health infrastructure, which can lead to delays from the time cases are first detected, their laboratory confirmation and finally public communication. In collaboration with the CMAEE2 lab, our project proposes a new method in the epidemic intelligence domain that is designed to discover knowledge in heterogenous web documents dealing with animal disease outbreaks. The proposed method consists of four stages: data acquisition, information retrieval (i.e. identification of relevant documents), information extraction (i.e. extraction of symptoms, locations, dates, diseases, affected animals, etc.), and evaluation by different epidemiology experts (Arsevska et al., 2014). 3.2 Information extraction from
experimental data Our joint work with the IATE3 lab and AgroParisTech4 deals with knowledge engineering issues regarding the extraction of experimental data from scientific papers to be subsequently reused in decision support systems. Experimental data can be represented by n-ary relations, which link a studied topic (e.g. food packaging, transformation process) with its features (e.g. oxygen permeability in packaging, biomass grinding). This knowledge is capitalized in an ontological and terminological resource (OTR). Part of this work consists of recognizing specialized terms (e.g. units of measures) that have many lexical variations in scientific documents in order to enrich an OTR (Berrahou et al., 2013).
1http://www.textmining.biz/agroNLP. html
2Joint research unit (JRU) regarding the control of exotic and emerging animal diseases – http://umr-cmaee. cirad.fr
3JRU in the area of agro-polymers and emerging technologies – http://umr-iate.cirad.fr 4http://www.agroparistech.fr Heterogenous data processing enables us to address several text-mining issues. Note that we integrated the knowledge of experts in the core of research applications summarized in Section 3.
In future work, we plan to investigate other techniques dealing with heterogeneous data, such as visual analytics approaches (Keim et al., 2008). to integrate declarative query constructs from the database community into MapReduce to allow greater data independence. 3
Relational DBMS Since it was developed by Edgar Codd in 1970, as presented in (Shuxin and Indrakshi, 2005), the relational database (RDBMS) has been the dominant model for database management. RDBMS is the basis for SQL, and is a type of database management system (DBMS) that is based on the relational model which stores data in the form of related tables, and manages and queries structured data. Since the RDBMSs focuse on extending the database system’s capabilities and its processing abilities, RDBMSs have become a predominant powerful choice for the storage of information in new databases because they are easier to understand and use. What makes it powerful, is that it is based on relation between data; because the possibility of viewing the database in many different ways since the RDBMS require few assumptions about how data is related or how it will be extracted from the database. So, an important feature of relational systems is that a single database can be spread across several tables which might be related by common database table columns.
RDBMS also provide relational operators to manipulate the data stored into the database tables.
However, as discussed in (Hammes et al., 2014), the lack of the RDBMS model resides in the complexity and the time spent to design and normalize an efficient database. This is due to the several design steps and rules, which must be properly applied such as Primary Keys, Foreign Keys, Normal Forms, Data Types, etc. Relational Databases have about forty years of production experience, so the main strength to point out is the maturity of RDBMSs. That ensure that most trails have been explored and functionality optimized. For the user side, he must have the competence of a database designer to effectively normalize and organize the database, plus a database administrator to maintain the inevitable technical issues that will arise after deployment.
A lot of work has been done to compare the MapReduce model with parallel relational databases, such as (Pavlo et al., 2009), where experiments are conducted to compare Hadoop MapReduce with two parallel DBMSs in order to evaluate both parallel DBMS and the MapReduce model in terms of performance and development complexity. The study showed that both databases did not outperformed Hadoop for user-defined function. Many applications are difficult to express in SQL, hence the remedy of the user-defined function. Thus, the efficiency of the RDBMSs is in regular database tasks, but the user-defined function presents the main ability lack of this DBMS type.
A proof of improvement of the RDBMS model comes with the introduction of the Object-Oriented Database Relational Model (ORDBMS). It aims to utilize the benefits of object oriented theory in order to satisfy the need for a more programmatic flexibility. The basic goal presented in (Sabàu, 2007) for the Object-relational database is to bridge the gap between relational databases and the object-oriented modeling techniques used in programming languages. The most notable research project in this field is Postgres (Berkeley University, Californie); Illustra and PostgreSQL are the two products tracing this research. 4
MapReduce-RDBMS: integrating paradigms In this section, the use of relational DBMS and MapReduce as complementary paradigms is considered.
It is important to pick the right database technology for the task at hand. Depending on what problem the organization is trying to solve, it will determine the technology that should be used.
Several comparative studies have been co ducted between MapReduce and parallel DBMS such as (Mchome, 2011) and (Pavlo et al., 2009).
MapReduce has been presented as a replacement for the Parallel DBMS. While each system has its strengths and its weaknesses. However, an integration of the two systems is needed, and as proposed in (Stonebraker et al., 2010), MapReduce can be seen as a complement to a RDBMS for analytical applications, because different problems require complex analysis capabilities provided by both technologies.
In this context, we propose a model that integrates the MapReduce model and a relational DBMS PostgreSQL presented in (Worsley and Drake, 2002), in a goal of queries optimization.
We suggest an OLAP queries process model in a goal of minimizing Input/Output costs in terms of the amount of data to manipulate, reading and writing throughout the execution process.
The basic idea behind our approach is based on the cost model to approve execution and selectivity of solutions based on the estimated cost of execution. To support the decision making process for analyzing data and extracting useful knowledge while minimizing costs, we propose to compare the estimates of the costs of running a query on Hadoop MapReduce compared to PostgreSQL to choose the least costly technology.
As the detailed analysis of the queries execution costs showed a gap mattering between both paradigms, hence the idea of the thorough analysis of the execution process of each query and the implied cost. So, to better control the cost difference between costs of Hadoop MapReduce versus PostgreSQL on each step of the query's execution process, we propose to dissect each query for a set of operations that demonstrates the process of executing the query and in order to control the different stages of the execution process of each query (decomposing the query to estimate the costs on each system for each individual operation). In this way, we can check the impact of the execution of each operation of a query on the overall cost and we can control the total cost of the query by controlling the partial cost of each operation in the information retrieval process. For this purpose, we suggest to provide a detailed execution plan for OLAP queries. This execution plan allows zooming on the sequence of steps of the process of executing a query. It details the various operations of the process highlighting the order of succession and dependence. In addition, it determines for each operation the amount of data involved and the dependence implemented in the succession of phases. These parameters will be needed to calculate the cost involved in each operation.
Having identified all operations performed during the query execution process the next step is then to calculate the cost implied in each operation independently, in both paradigms PostreSQL and MapReduce with the aim of controlling the estimated costs difference according to the operations as well as the total cost of query execution. At this stage we consider each operation independently to calculate an estimate of its cost execution on PostreSQL on one hand then on MapReduce on the other hand. For this goal, we propose to relay on a cost model for each system PostgreSQL and Hadoop MapReduce in order to estimate the I/O cost of each operation execution on both system independently.
We aim to estimate how expensive it is to preprocess each operation of each query on both systems. Therefore, controlling the cost implied by each operation as well as its influence on the total cost of the query, allows the control of the cost of each query to support the decision making process and the selectivity of the proposed solutions based on the criterion of cost minimization.
Based on a sample workload of OLAP queries, and having identifying the cost of each operation for each query executed on both systems independently, the results analysis can be useful to deduct a generalized smart model that integrates the two paradigms to process the OLAP queries in a cost minimization way. 5 Given the exploding data problem, the world of databases has evolved which aimed to escape the limitations of data processing and analysis. There has been a significant amount of work during the last two decades related to the needs of new supporting technologies for data processing and knowledge management, challenged by the rise of data generation and data structure diversity.
In this paper, we have investigated the MapReduce model in one hand, then the Relational DBMS technology in the other hand, in order to present strengths and weaknesses of each paradigm.
Although MapReduce was designed to cope with large amounts of unstructured data, there will be advantages in exploiting it in structured data processing. In this fashion, we have proposed in this paper a new OLAP queries process model integrating an RDBMS with the MapReduce framework in a goal of minimizing Input/Output costs in terms of the amount of data to manipulate, reading and writing throughout the execution process.
Combining MapReduce and RDBMS technologies has the potential to create very powerful systems. For this reason, we plan to investigate other types of integration for different applications.
Hammes D., Medero, H. and Mitchell H. 2014. Comparison of NoSQL and SQL Databases in the Cloud.
Southern Association for Information Systems (SAIS) Proceedings. Paper 12.
McClean A., Conceicao RC., and O'Halloran M. 2013.
A Comparison of MapReduce and Parallel Database Management Systems. ICONS 2013, The Eighth International Conference on Systems: 64-68.
Mchome M.L. 2011. Comparison study between
MapReduce (MR) and parallel data management systems (DBMs) in large scale data anlysis. Honors
Projects Macalester College.
Nykiel T., Potamias M., Mishra C., Kollios G. and
Koudas N. 2010. Mrshare: sharing across multiple queries in mapreduce. Proceedings of the VLDB Endowment ,volume 3 (1-2). 494-505.
Ordonez C. 2013. Can we analyze big data inside a
DBMS?. Proceedings of the sixteenth international workshop on Data warehousing and OLAP, ACM.
85-92.
Olston C., Reed B., Srivastava U., Kumar R. and Tomkins A. 2008. Pig latin: a not-so-foreign language for data processing. Proceedings of the 2008 ACM SIGMOD international conference on Management of data , ACM. 1099-1110.
Pavlo A., Rasin A., Madden S., Stonebraker M.,
DeWitt D., Paulson E., Shrinivas L. and Abadi D.J. 2009. A comparison of approaches to large scale data analysis. Proceedings of the 2009 ACM SIGMOD International Conference on Management of data, ACM. 165-178.
Shuxin Y. and Indrakshi R. 2005. Relational database operations modeling with UML. Proceedings of the 19th International Conference on Advanced Information Networking and Applications. 927-932.
Sabàu G. 2007. Comparison of RDBMS, OODBMS
and ORDBMS. Informatica Economic.
Stonebraker M., Abadi D., DeWitt D.J., Madden S.,
Paulson E., Pavlo A. and Rasin A. 2010. Mapreduce and parallel dbmss : friends or foes ?. Communications of the ACM, volume 53 (1). 64-71.
Wang G. and Chan CY. 2013. Multi-Query Optimization in MapReduce Framework. Proceedings of the VLDB Endowment, 40th International Conference on Very Large Data Bases, volume 7 (3).
Worsley J.C. and Drake J.D. 2002. Practical PostgreSQL. O'Reilly and Associates Inc Performance of Alternating Least Squares in a distributed approach using
GraphLab and MapReduce Elizabeth Veronica Vera Cervantes, Laura Vanessa Cruz Quispe, Jose´ Eduardo Ochoa Luna National University of San Agustin
Arequipa, Peru elizavvc@gmail.com, lcruzq@unsa.edu.pe, eduardo.ol@gmail.com Automated recommendation systems have been increasingly adopted by companies that aim to draw people attention about products and services on Internet. In this sense, development of distributed model abstractions such as MapReduce and GraphLab has brought new possibilities for recommendation research tasks due to allow us to perform Big Data analysis. Thus, this paper investigates the suitability of these two approaches for massive recommendation. In order to do so, the Alternating Least Squares (ALS), which is a Collaborative Filtering algorithm, has been tested using recommendation benchmark datasets. Results on RMSE show a preliminary comparative performance analysis. 1 Introduction Data on the Internet is increasing, e-commerce sites, blogs and social networks spread the word about new products and services everyday. This social media information overwhelms any user, who has a given profile and therefore could not be interested in most of these offers (Koren et al., 2009).
In this sense, recommendation systems have gained momentum, because they “filter” products and services for users according to behavior patterns. Traditional approaches for automated recommendation range from Content-Based, Collaborative Filtering and Deep Learning systems (Adomavicius, 2005; Shi et al., 2014). However, to handle the current amount of available data we need to resort to frameworks for large-scale data processing.
Recently, the machine learning community has been increasingly interested in the task of managing Big Data with parallelism (Zhou et al., 2008; De Pessemier et al., 2011; Xianfeng Yang, 2014).
However, parallel algorithms are extremely challenging and traditional approaches, despite of being powerful like MPI, rely on low levels of abstraction. On the other hand, distributed models such as GraphLab(Low et al., 2012) and MapReduce(Xiao and Xiao, 2014; Dean and Ghemawat, 2008) foster high levels of abstraction and, therefore, they are more intuitive. The aim of this paper is to investigate whether these distributed models are suitable for recommendation tasks. In order to do so we evaluate the Alternating Least Squares (ALS) algorithm, a parallel collaborative filtering approach(Koren et al., 2009; Schelter et al., 2013), in both GraphLab and MapReduce frameworks. We evaluate the performance on the MovieLens and Netflix datasets. According to preliminary results, GraphLab outperforms MapReduce in RMSE, when Lambda, iterations number and latent factor parameters are considered. Conversely, MapReduce gets a better execution time than GraphLab using the same parameters in MoveLens dataset. The paper is organized as follows. In Section 2, related work is described.
Background is given in Section 3. Our proposal is showed in Section 4. Preliminary results are depicted in section 5. Finally Section 5, concludes the paper. 2 Several distributed platforms have been used for studying performance of machine learning algorithms, for instance, a Matrix Factorization based on collaborative filtering over MapReduce model was proposed in (Xianfeng Yang, 2014; De Pessemier et al., 2011). In Low et al. (2012), some advantages and disadvantages of using GraphLab and MapReduce were described.
For instance, MapReduce fails when there are computational dependencies on data, but it can be used to extract features from a massive collection. In addition, MapReduce is targeted for large data centers, it is optimized for node-failure and diskcentric parallelism. Conversely, In GraphLab it is assumed that processors do not fail, and all data is stored in shared memory.
In Low et al. (2012), the Alternating Least Squares (ALS) algorithm was implemented over several platforms: GraphLab, Hadoop/MapReduce and MPI. Comparison results show that applications created using GraphLab outperformed equivalent Hadoop/MapReduce implementations by 20-60 times(Xianfeng Yang, 2014) .
Our work is most related to Low et al. (2012), but we focus on the evaluation of different configurations of ALS algorithm over GraphLab and MapReduce. Thus, we aim at obtaining optimal parameters that allow us to improve algorithm performance. Moreover, comparisons were based on RMSE and time execution values. The parameters considered are: • Lambda, which is the regularization parame
ter in ALS 3 3.1 • The number of latent factors • The number of iterations
Background A recommendation system aims at showing items of interest to a user, considering the context of where the items are being shown and to whom they are being shown (Alag, 2008).
Figure 1, depicts inputs and outputs of a common recommendation system. • Content-based Recommendation: Items similar to the ones he/she has preferred in the past, are recommended to the user. • Collaborative Recommendation: Items that people with similar tastes and preferences liked in the past, are recommended to the user.
– Collaborative Deep Learning: It is a recent kind of collaborative filtering using deep learning models Wang et al.
(2014). • Hybrid Approach: Recommendations are made using a combination of Content-based and Collaborative Recommendation methods.
Alternating Least Squares (ALS) Alternating Least Squares (Low et al., 2012; Zhou et al., 2008; Koren et al., 2009) is an algorithm within the collaborative filtering paradigm. Input of ALS (in Figure 2) is a sparse user by items matrix R containing the rating of each user. The algorithm iteratively computes a low-rank matrix factorization R = U × V where U and V are d dimensional matrices. The loss function is defined as the squared error(Zhou et al., 2008), where the learning objective is to minimize the sum of the squared errors 1 between values predicted and real values of rantings.
(Uˆ , Vˆ ) = argmin X i, j ∈ R(rij − viT uj )2 (1)
U,V Complexity and cost depend on the magnitude of the hidden variables d.
U=Users
R d
C=Courses d
In Adomavicius (2005) three approaches for building a recommendation system are presented:
The ALS algorithm is computationally expensive, every iteration runs on O(d−1[N r + (m + n)r2] + r3), where m is length of items, and n is length of users (Schelter et al., 2013; Gemulla et al., 2011).
Alternating Least Squares on GraphLab According to (Low et al., 2012; Gonzalez et al., 2012) ALS in GraphLab is implemented by using a bipartite two colorable graph and a chromatic synchronous engine with an edge consistency model for serializability.
Each vertex of the graph has a latent factor attached, that denotes a user or an item. Thus, they are linked to a column or a row in the matrix of ratings R. Each edge of the graph contains entry data (rating values), and the most recent error estimated by the algorithm. The goal of ALS algorithm is to discover values of latent parameters, such that non-zero entries in R can be predicted by the dot product of the row and column latent factor. ALS algorithm for GraphLab is implemented in the Gather-Apply-Scatter abstraction. ALS update considers adjacent vertices as X values and edges as observed y values, and then updates the current vertex value as a weight w:
y = X ∗ w + noise w = inv(X0 ∗ X) ∗ (X0 ∗ y) In the Gather-Apply-Scatter model, the update is done as follows: • Gather: it returns the tuple (X0 ∗ X, X0 ∗ y) • Apply: it solves inv(X0 ∗ X) ∗ (X0 ∗ y) • Scatter: it schedules the update of adjacent vertices if this vertex has changed and the edge is not well predicted. 3.4
Alternating Least Squares on
MapReduce In Xianfeng Yang (2014; Zhou et al. (2008) MapReduce implementation is comprised by four tasks as shown in Figure 3. Each item in dataset is denoted as a triple (u, j, r). u denotes user, j is the label of item and r denotes corresponding rating. In the U-Update step, item matrix V is used as input and is sent to cluster nodes. Then, training rating R is used to compute user matrix U , including inputs as lambda parameter λ to regularization, number of latent factors. V-Update does (2) (3) that is accomplished using the following equation: 4 the same as U-Update step, but its input is not an item matrix. On the contrary, it is a user matrix computed in U-Update step. Once U and V are learned, we can compute RMSE values using test dataset and estimated rating rˆ. So the Parallel ALS algorithm with Weighted--Regularization is as follows (Zhou et al., 2008): The objective function in Algorithm 1 Alternating Least Square(ALS) with algorithm 1: Initialize V with random values between 0
and 1 2: Hold V constants, and solve U by minimizing
the objective function. 3: Hold U constants, and solve M by minimiz
ing the objective function. 4: repeat from step 2 and 3 until objective func
tion converge. 1 is obtained from equation 4, which is just linear regression with lambda regularization(λ), to avoid overfits it penalize large parameters.
In this paper we evaluate several parameter configurations (lambda, number of latent factor, number of iterations) for ALS algorithm over GraphLab and MapReduce. Our aim is to obtain the best performance, over clusters of two and four machines, for the Movielens Dataset, and NetFlix Dataset (further details will be given in the next section).
We evaluate performance according to RMSE and execution time values.
In order to implement ALS algorithm under the MapReduce Paradigm, the Mahout 1 API has been used. ALS algorithm for MapReduce (Zhou et al., 2008) is shown in 3. User and movie factors have been computed using equation 4. where nui and nvj are the numbers of ratings of user i and item j respectively. When objective function showed in equation 4 does not change after further iterations, we attain the final step. Output is the predicted rating for each user/item pair.
1http://mahout.apache.org/ f (U, V ) =
X (rij − uiT vj )2 + λ(X nuik ui k2 + i,j∈I i Trainingratingmatrix, itemmatrix V-Update
Trainingratingmatrix, usermatrix U-Update
Testratingmatrix, itemmatrix,usermatrix RMSECalculation
Bestregularizationfactor ƛ, itemmatrix,usermatrix Prediction estimatedratingsfor user/itempairs
In order to evaluate ALS algorithm under GraphLab, the GraphLab API (Low et al., 2010) has been used. ALS algorithm for GraphLab (Low et al., 2012) is shown in Figure 4. User and movie factors have been computed using equation 5.
f [i] = argmin w∈Rd j∈Neighbors(i)
X
(rij − wT f [j]) (5) t)(r U s o c a fr e s U
C o u r s e f a c t()r o s
C
Matrix factorization of ALS using MovieLens is a Web collaborative site that manages a recommender system for movies. This recommender system is based on a collaborative filtering algorithm developed by the GroupLens research group. The dataset is comprised by 6040 users, 3952 items and 100209 ratings for training.
The data structure is: user, item, rating. 4.2
Netflix Dataset We are using the Small Netflix Dataset. It is also a data-set for movie recommendation, it has 95526 users, 3561 items and 3298163 ratings. The structure of the data-set is: user, item, rating. Setup of the GraphLab cluster is as follows. Two machines, one working as the master and the other as the worker node. The master machine operating system is Ubuntu 14.04, and its processor is Intel Core i3 CPU M 330@2.13GHzx4. The worker machine operating system is Ubuntu 13.10 of 64-bit, and its processor is Intel Core i3-2350M @2.30GHzx4. The cluster was configured using MPI(Message Passing Interface). The setup is as follows. Four machines, three worker nodes and one master. The master machine operating system is Ubuntu 13.10 of 64-bit, and its processor is Intel Core i3-2350M @2.30GHzx4.
Table.1 shows the congfiuration of the worker machines.
The cluster was configured using Hadoop, and Machine Operating System Processor 1 Ubuntu 14.04 Intel Core i3 CPU M 330@2.13GHzx4 2 Ubuntu 13.10 Intel Core i3-2350M @2.30GHzx4 3 Ubuntu 14.04 Intel Core i7-4700MQ @2.40GHzx8 the HDFS(Hadoop Distributed File System). The ALS(Alternating Least Squares) algorithm implementation was taken from Mahout Library. This section shows experimental results conducted on MovieLens data set aforementioned.
Experimental setting parameters are described in
Parameters Lambda # Latent factors # Iterations
Value 0.01 - 0.09 10-50 2-30 Table 2: Parameters used for ALS algorithms
In Figure 5a, RMSE values for MapReduce do not change even if we increase the number of latent factors thus, RMSE values on MapReduce are independent on the number of latent factors. RMSE for MapReduce converges around 0.95. Conversely, RMSE values for GraphLab decreases while the number of latent factors increases. When the number of latent factors was 50, RMSE value reaches around 0.25. However, GraphLab spends more time than MapReduce, Figure 5b depicts MapReduce times almost as an horizontal line for MovieLens dataset, the line of execution time for Netflix dataset is much steeper.
Between Graphlab and MapReduce lines representing Movilens dataset execution, Graphlab line is more pronounced.
Figure 6a depicts GraphLab and MapReduce performance according the Lambda parameter.
While Lambda increases, RMSE decreases accordingly, i.e., if a greater value of Lambda is used then algorithm accuracy tends to be better. We also notice that Graphlab has lower values of RMSE compared to MapReduce. GraphLab RMSE values are around 0.5, and MapReduce RMSE values are around 1. Figure 6b illustrates a better execution time of MapReduce compare to GraphLab over Movielens dataset. However, now the execution time for GraphLab decreases, while the value of Lambda increases. Figure 6b also shows that It takes longer to process the data from netflix than Movielens.
In Figure 7a we notice that the value of RMSE is almost invariant to the increase of iterations for MapReduce execution, given that the number of iterations are small,nevertheless we notice clearly that RMSE value for GraphLab decreases as the number of iterations increases. RMSE value for Graphlab converges around 0.55. Figure 7b shows that MapReduce execution time over Movielens dataset is good, however it increases a lot for Netflix dataset. Graphlab execution time increases as the number of iterations grows. 1 0.9 0.8 0.7 ESM 0.6 R 0.5 0.4 0.3
Netflix(M) MovieLens(M)
MovieLens(G) 15 20 25 30 35 40
Number of Latent Factors (b) Number of latent factors Vs. Time Figure 5: Performance of MapReduce and GraphLab when number of features in ALS algorithms is increased.
111...321 MMoovvNiieeeLLteeflnnixss(((MMG))) 1 ES 0.9 RM 0.8 0.7 0.6 0.5 0.40.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09
Lambda (a) Lambda Vs. RMSE 0.004..545 MMoovvNiieeeLLteeflnnixss(((MMG))) )r 0.35 s 0.3 u oh 0.25 ( e iTm 0.2 0.15 0.1 0.05 00.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09
Lambda (b) Lambda Vs. Time
Netflix(M) MovieLens(M)
MovieLens(G) 15
20 Number of Iterations 25
30 (a) Number of Iterations Vs. RMSE 15
20 Number of Iterations 25
30 (b) Number of Iterations Vs. RMSE We evaluated the Alternating Least Squares (ALS) algorithm, a parallel collaborative filtering in both GraphLab and MapReduce frameworks.
Experiments were run over the MovieLens and Netflix datasets. The RMSE between MapReduce execution in NetFlix dataset and Movielens dataset in all the experiments was similar, but the execution time was longer in Netflix dataset.
Looking at the executions over Moviliens dataset, we can say, that even though GraphLab only ran in two machines and MapReduce in 4 machines, the first one outperformed the second one in RMSE.
Considering lambda value variation, Figure.6a,the number of iterations Figure.7a, and the number of latent factors Figure.5a, GraphLab performed better (RMSE) than MapReduce. In all previous three cases MapReduce was faster than GraphLab, obviously by the difference between the number of machines in their configuration.
Thus, when scalability and distribution are evaluated, MapReduce performs better, because ALS does not require data dependency for computing.
Moreover, it took less execution time when more latent factors were added. In this work we only used two nodes, however GraphLab demonstrated best results with few nodes.
In conclusion, GraphLab performed better when RMSE was considered but, there are open issues with shared-memory. GraphLab is also better for computing recommendations in real time.
However, for more sophisticated computations MapReduce performs better so far as to an offline environment and all data is used.
References Tuzhilin A. Adomavicius, G. 2005. Toward the next generation of recommender systems: a survey of the state-of-the-art and possible extensions. Knowledge and Data Engineering, IEEE Transactions on, 17:734 – 749.
Satnam Alag. 2008. Collective Intelligence in Action.
Toon De Pessemier, Kris Vanhecke, Simon Dooms, and Luc Martens. 2011. Content-based recommendation algorithms on the hadoop mapreduce framework. In 7th international conference on Web Information Systems and Technologies, Proceedings, pages 237–240. Ghent University, Department of Information technology.
Jeffrey Dean and Sanjay Ghemawat. 2008. Mapreduce: Simplified data processing on large clusters.
Commun. ACM, 51(1):107–113, January.
Rainer Gemulla, Erik Nijkamp, Peter J Haas, and Yannis Sismanis. 2011. Large-scale matrix factorization with distributed stochastic gradient descent.
In Proceedings of the 17th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 69–77. ACM.
Joseph E. Gonzalez, Yucheng Low, Haijie Gu, Danny
Bickson, and Carlos Guestrin. 2012. Powergraph: Distributed graph-parallel computation on natural graphs. In Proceedings of the 10th USENIX Conference on Operating Systems Design and Implementation, OSDI’12, pages 17–30, Berkeley, CA, USA.
USENIX Association.
Y Koren, R Bell, and C Volinsky. 2009. Matrix factorization techniques for recommender systems. Computer, 42(8):30–37.
Yucheng Low, Joseph Gonzalez, Aapo Kyrola, Danny
Bickson, Carlos Guestrin, and Joseph M. Hellerstein. 2010. Graphlab: A new parallel framework for machine learning. In Conference on Uncertainty in Artificial Intelligence (UAI) , July.
Yucheng Low, Danny Bickson, Joseph Gonzalez, Carlos Guestrin, Aapo Kyrola, and Joseph M Hellerstein. 2012. Distributed graphlab: a framework for machine learning and data mining in the cloud. Proceedings of the VLDB Endowment, 5(8):716–727.
Sebastian Schelter, Christoph Boden, Martin Schenck,
Alexander Alexandrov, and Volker Markl. 2013.
Distributed matrix factorization with mapreduce using a series of broadcast-joins. In Proceedings of the 7th ACM Conference on Recommender Systems, RecSys ’13, pages 281–284, New York, NY, USA.
ACM.
Yue Shi, Martha Larson, and Alan Hanjalic. 2014.
Collaborative filtering beyond the user-item matrix: A survey of the state of the art and future challenges.
ACM Comput. Surv., 47(1):3:1–3:45, May.
Hao Wang, Naiyan Wang, and Dit-Yan Yeung. 2014.
Collaborative deep learning for recommender systems. CoRR, abs/1409.2944.
Pengfei Liu Xianfeng Yang. 2014. Collaborative filtering recommendation using matrix factorization: A mapreduce implementation. International Journal of Grid and Distributed Computing.
Zhifeng Xiao and Yang Xiao. 2014. Achieving accountable mapreduce in cloud computing. Future
Gener. Comput. Syst., 30:1–13, January.
Yunhong Zhou, Dennis Wilkinson, Robert Schreiber, and Rong Pan. 2008. Large-scale parallel collaborative filtering for the netflix prize. In Proceedings of the 4th International Conference on Algorithmic Aspects in Information and Management, AAIM ’08, pages 337–348, Berlin, Heidelberg. SpringerVerlag.
Data Modeling for NoSQL Document-Oriented Databases Harley Vera, Wagner Boaventura, Maristela Holanda, Valeria Guimar a˜es, Fernanda Hondo Department of Computer Science
University of Bras´ılia
Bras´ılia, Brasil. {harleyve,wagnerbf}@gmail.com, mholanda@cic.unb.br, {valeriaguimaraes, fernandahondo}@hotmail.com In database technologies, some of the new issues increasingly debated are non-conventional applications, including NoSQL (Not only SQL) databases, which were initially created in response to the needs for better scalability, lower latency and higher flexibility in an era of bigdata and cloud computing. These non-functional aspects are the main reason for using NoSQL database. However, currently there are no systematic studies on data modeling for NoSQL databases, especially the document-oriented ones.
Therefore, this article proposes a NoSQL data modeling standard in the form of ER diagrams, introducing modeling techniques that can be used on documentoriented databases. On the other hand the purpose of this article is not structure the data using the model proposed, but it does helping with the visualization of data. In addition, to validate the proposed model, a study case was implemented using genomic data. 1 Introduction Huge amounts of data are produced daily. They are generated by smart phones, social networks, banks transactions, machines measured by sensors are part of Internet of Things provide information that is growing exponentially. The management of this data is currently performed in most cases by relational databases that provide centralized control of data, redundancy control and elimination of inconsistencies (Elmasri and Navathe, 2010); but, some of these factors restrict the use of alternative database models. Consequently, certain limiting factors have led to alternative models of databases in these scenarios. Primarily, motivated by the issue of system scalability, a new generation of databases, known as NoSQL, is gaining strength and space in information systems. The NoSQL databases emerged in the mid-90s, from a database solution that did not provide an SQL interface. Later, the term came to represent solution that promote an alternative to the Relational Model, becoming an abbreviation for Not Only SQL.
The purpose, therefore, of NoSQL solutions is not to replace the Relational Model as a whole, but only in cases in which there is a need for scalability and bigdata. In the recent years, a variety of NoSQL databases has been developed mainly by practitioners looking to fit their specific requirements regarding scalability performance, maintenance and feature-set. Subsequently, there have been various approaches to classify NoSQL databases, each with different categories and subcategories, such as key-value stores, columnoriented and graph databases, oriented-document.
MongoDB
Nonetheless, data modeling still has an important role to play in NoSQL environments. The data modeling process (Elmasri and Navathe, 2010) involves the creation of a diagram that represents the meaning of the data and the relationship between the data elements. Thus, understanding is a fundamental aspect of data modeling (R. F. Lans, 2008), and a pattern for this kind of representation has few contributions for NoSQL databases.
Addressing this issue, this article proposes a standard for NoSQL data modeling. This proposal uses NoSQL document-oriented databases, aiming to introduce modeling techniques that can be used on databases with document features.
The remainder of the paper is organized as follows: Section II presents related works. Section III explores the concepts of modeling for NoSQL databases based on documents, introducing the different types of relationships and associations.
Section IV shows the proposal model in the context of NoSQL databases based on documents.
Section V presents the study case to validate the proposal model. Finally in Section VI, presents the conclusion of the research and future works. 2 Katsov (H. Scalable, 2015) presents a study of techniques and patterns for data modeling using different categories of NoSQL databases. However, the approach is generic and does not define a specific modeling engine to each database.
Arora and Aggarwal (R. Arora and R. Aggarwal, 2013) propose a data modeling, but restricted to MongoDB document database, describing a UML Diagram Class and JSON format to represent the documents.
Similarly, Banker (K. Banker, 2011) provides some ideas of data modeling, but limited to MongoDB database and always referring to JSON (D.
Crockford, 2006) format as a modeling solution.
Kaur and Rani
In the case of document-oriented propose a data modeling restricted to MongoDB document database, describing the data model by UML diagram class to represent documents. 3
Data Modeling For
Document-Oriented Database
An important step in database implementation is
the data modeling, because it facilitates the
understanding of the project through key features
that can prevent programming and operation
errors. For relational databases, the data
modeling uses the Entity-Relationship Model (Elmasri
and Navathe, 2010). For NoSQL, it depends on
the database category. The focus of this article is
NoSQL document-oriented databases, where the
data format of these documents can be JSON,
BSON, or XML
Basically, the documents are stored in collections. A parallel is made with relational databases, the equivalent for a collection is the record (tuple) and for a document it is the relation (table).
Documents can store completely different sets of attributes, and can be mapped directly to a file format that can be easily manipulated by a programming language. However, it is difficult to abstract the modeling of documents for the entity relationship model (R. F. Lans, 2008).
Modeling Paradigm for document-oriented Database The relational model designed for SQL has some important features such as integrity, consistency, type validation, transactional guarantees, schemes and referential integrity. However, some applications do not need all of these features. The elimination of these resources has an important influence on the performance and scalability of data storage, bringing new meaning to data modeling.
Document-oriented databases have some
significant improvements, e.g., index management by
the database itself, flexible layouts and advanced
indexed search engines (H. Scalable, 2015). By
associating these improvements (some being
denormalization and aggregation) to the basic
principles of data modeling in NoSQL, it is
possible to identify some generic modeling standards
associated to document-oriented databases.
Analyzing the documentation of the main
documentoriented databases, MongoDB
Proposal For Document-Oriented
Databases Viewing Unlike traditional relational databases that have a simple form in the disposition in rows and columns, a document-oriented database stores information in text format, which consists of collections of records organized in key-value concept, ie, for each value represented a name (or label) is assigned, which describes its meaning. This storage model is known as JSON object, and the objects are composed of multiple name/value pairs for arrays, and other objects.
In this scenario, the number of objects in a database increases the abstraction complexity of the logical relationship between the stored information, especially when objects have references to other objects. Currently, there is a lack of solutions to conceptually represent those associated with a NoSQL document-oriented database. As described in (R. Arora and R. Aggarwal, 2013), there is no standard to represent this kind of object modeling, several diferent manners of modeling may arise, depending on each data administrator’s understanding, which makes learning difficult for those who need to read the database model.
Therefore, this section proposes a standard for document-oriented database viewing. Our proposal has some properties, considering the conceptual representation modeling type, such as: • Ensuring a single way of modeling for the several NoSQL document-oriented databases. • Simplifying and facilitating the understanding of a document-oriented database through its conceptual model, leveraging the abstraction and making the correct decisions about the data storage. • Providing an accurate, unambiguous and concise pattern, so that database administrators have substantial gains in abstraction, understanding. • Presenting different types of relationships between collections defined as References and
Embedded documents. • Assisting the recognition and arrangement of the objects, as well as its features and relationships with other objects.
The following subsections present the concepts and graphing to build a conceptual model for NoSQL document-oriented databases. Before starting the discussion about the approach of each type of the conceptual modeling representation, it is important to highlight some basic concepts about objects and relationships in a document-oriented database: • A document (or object) describes a set of attributes that have their properties organized in a key-value structure. • Information contained in an document is described by the identifier (key) and the value associated with the key. • Different types of relationships between documents are defined as References and Embedded Documents • Because NoSQL is a non-relational data database, the concepts of normalization, do not apply. • Some concepts of relationships between objects are similar to ER modeling, such as cardinality (one-to-one, one-to-many, many-tomany). The proposed solution for a conceptual modeling to the NoSQL document-oriented databases has two basic concepts: Document and Collections.
As noted previously, a document is usually represented by the structure of a JSON object, and as many fields as needed may be added to the document. For this solution, a document and a collection of documents is represented by Figure 3.
Embedded Documents 1..N A one-to-many relationship in embedded documents is represented by the Figure 5. This is the case when the notation to represent the cardinality is the same used in UML (F. Booch et al., 2005) and is placed in the upper right corner of the embedded documents. According to the cardinality one-to-many the larger document has embedded multiple documents within it.
The following section presents the definitions of relationship types and degrees for the objects features.
Embedded Documents 1..1 This section proposes a model that represents the one-to-one relationship for documents embedded in another document. In this case, the proposal is to use the representation of an individual Document within another element that represents a Document. In Figure 4, cardinality is also suggested to specify the one-to-one relationship type. A many-to-many relationship in embedded documents is represented by the Figure 6. According to the cardinality many-to-many the larger document has a many to many relationship with the embedded document. The representation of the cardinality is the same used in UML (F. Booch et al., 2005). A document can reference another, and in this case, one must use an arrow directed to the referenced document, as shown in Figure 7. One can see that the directed arrow makes the left document references to the right document. Furthermore, the cardinality of the relationship should be specified above the arrow. The notation of cardinality is based on UML (F. Booch et al., 2005). In NoSQL, a document can reference multiple documents. To represent this relationship one should use an arrow directed to the referenced documents, as shown in Figure 8. The left document references multiple documents on the right side, by the directed arrow. Furthermore, the cardinality of the relationship is represented by the notation ”1..N” as in UML (F. Booch et al., 2005). 5 In order to evaluate our proposal, part of the workflow described in (J. C. Marioni et al., 2014) was used. This workflow aimed at identify and comparing expression levels of human kidney and liver RNA samples sequenced by Illumina. The workflow was designed in three phases (Figure. 10): • Filtering: all the sequenced transcripts were filtered, generating new files with good quality sequences. • Alignment: transcripts were mapped to the
human genome used as reference. • Statistical Analysis: a sort process was first executed, followed by a statistical analysis with the mapped transcripts to discover which genes are mostly expressed both in kidney and liver samples. To represent this relationship a bidirectional arrow is used between reference documents, as shown in Figure 9. The left document references multiple documents on the right side and the right document references multiple documents on the left side. Furthermore, the cardinality of the relationship is represented by the notation ”N..N” as in UML (F. Booch et al., 2005).
After analyzing the previously mentioned concepts, we have chosen to create a collection of documents for each PROV-DM type used to create a graph node. We also defined a collection for genomic documents (raw data). The reference relationship approach was chosen to connect all PROV-DM components, complementary information of PROV-DM and genomic documents. Based on (R. de Paula et al., 2013) we defined the documents and the attributes. A set of minimum information related to each one of these entities.
Figure. 11 shows our document based data representation, explained as follows:
• Project: stores different experiments of one agent. Attributes: Id, name, description, coordinator, start date, end date and observation. • FASTAQ: files used or generated in the activ
ity; Attributes: Id, filename and description. • Activity: represents the execution of a program; Attributes: Id, name, program, version program, command line, function, start date, end date, account ID, used (name FASTQ, local, size), wasGenerateBy (name FASTQ, local, size), and wasAssociatedWith (Agent name). • Account: represents the performance of an experiment; Attributes: Id, name, description, execution place, star date, end date, observation, version and version date and project Id. • Agent: represents the person responsible for a program or a phase in the workflow. Attributes: Id, name, login and password. 5.1 Implementation In this case study, we have considered the MongoDB NoSQL database to store provenance and data files. The primary motivation for this choice was MongoDB’s ability to manipulate large volumes of data. MongoDB is an open-source Document-Oriented database designed to store large amounts of data from multiple servers.
It uses JSON- style documents with dynamic
schemas. The number of fields, content and size
of the document can differ from one document to
another. In practice, however, the documents in
a collection share a similar structure
MongoDB documents have a maximum size of
16MB. This feature is important to ensure that
a single document cannot use excessive amounts
of RAM. In order to store files larger than the
maximum size, MongoDB provides a GridFS API
GridFS uses two collections to store the data: fs.files collections, containing metadata about files, and fs.chunks collections, which store the actual 256k data chunks. The collections FS.file contains the name of the FASTQ file. Thus, it was possible to implement the relationship between MongoDB Collection Activity using Reference Document. In other words, we implemented the connection between Level 1 and Level 2 through the File Name attribute that was present in fileprovenance.files and Activity Collection. Figure. 12 illustrates this particular implementation. In contrast to relational database management systems, NoSQL databases are designed to be schemaless and flexible. Therefore, the challenge of this work was to introduce a data modeling standard for NoSQL document-oriented databases, in contrast to the original idea for NoSQL databases.
The objective was to build compact, clear and intuitive diagrams for conceptual data modeling for NoSQL databases. While the current studies propose generic techniques and do not define a specific modeling engine to NoSQL database, our idea was to present a graphical model for any NoSQL document-oriented database. Moreover, while other studies describe techniques based on UML Diagram Class and JSON format as a modeling solution, we have a new approach to solve the conceptual data modeling issue for NoSQL document-oriented databases.
Future work includes: verifying our model for other NoSQL database classifications, such as key-value and column. R. Elmasri and S. Navathe. 2010. Fundamentals of
Database Systems. Pearson Addison Wesley.
MongoDB. 2015. Document database. [Online]
Available: http://www.mongodb.org/ [Retrieved:April, 15].
J. Partner, A. Vukotic, and N. Watt. 2013. Neo4j in
Action, O’Reilly Media.
D. Borthakur et al. 2011. Apache hadoop goes realtime at facebook, in Proceedings of the 2011 ACM SIGMOD International Conference on Management of data. ACM, 2011, pp. 1071–1080.
F. Chang et al. 2008. “Bigtable: A distributed storage system for structured data, ACM Transactions on Computer Systems (TOCS), vol. 26, no. 2, 2008, p.
4.
R. F. Lans. 2008. Introduction to SQL: mastering the relational database language, Addison-Wesley Professional.
H. Scalable. 2015. Nosql data modeling techniques. [Online] Available: http://highlyscalable. wordpress.com/2012/03/01/ nosql-data-modeling-techniques/ [Retrieved:April, 15].
R. Arora and R. Aggarwal, 2013. Modeling and querying data in mongodb, International Journal of Scientific and Engineering Research (IJSER 2013), vol.
4, no. 7, Jul. 2013, pp. 141–144.
K. Banker, 2011. MongoDB in action, Manning Pub
lications Co.
D. Crockford, 2006. RFC 4627 (Informational) The application json Media Type for JavaScript Object Notation (JSON), IETF (Internet Engineering Task Force) G. Booch, J. Rumbaugh, and I. Jacobson, 2005. The unified modeling language user guide. , Pearson Education India.
J. C. Marioni, C. E. Mason, S. M. Mane, M. Stephens, and Y. Gilad, 2014. RNA-SEQ: An assessment of technical reproducibility and comparison with gene expression arrays, Genome Research, vol. 18, no. 9, pp. 1509–1517.
R. de Paula, M. Holanda, L. SA Gomes, S. Lifschitz and M. E. MT. Walter, 2013. Provenance in bioinformatics workflows. , BMC Bioinformatics 14 (Suppl 11):S6.
Hipi, as alternative for satellite images processing Wilder Nina Choquehuayta
UNSA / Arequipa wninac@unsa.edu.pe
Rene´ Cruz Mu n˜oz
UNSA / Arequipa rcruzm@unsa.edu.pe
Juber Serrano Cervantes
UNSA / Arequipa jserranoc@unsa.edu.pe Alvaro Mamani Aliaga
UNSA / Arequipa amamani@unsa.edu.pe
Pablo Yanyachi
UNSA / Arequipa raulpab@unsa.edu.pe
Yessenia Yari
UNSA / Arequipa yyari@unsa.edu.pe These days, in different fields of both industry and academia, large amounts of data is generated. The use of several frameworks with different techniques is essential , for processing and extraction of data. In the remote sensing field, large volumes of data are generated (satellite images) over short periods of time. Information systems for processing these kind of images were not designed with scalable features. In this paper, we present an extension of the HIPI framework (Hadoop Image Processing Interface) for processing satellite image formats.
KeyWords: Big Data, Remote Sensing, Hadoop, HIPI, MapReduce, Satellite Images 1 Introduction The field of remote sensing is helpful in different areas of both industry and academia, because it uses images of the earth’s surface that are acquired from different sources like antennas and satellites, which provide increasingly better image resolution as technology advances. Nowadays, several open source frameworks are available to process big data, such as Hadoop (Shvachko et al., 2010), H2O (0xdata:H2O, 2015), Spark (Apache:Spark, 2015), etc., which are used for distributed and parallel processing of large volumes of data. HIPI (Hadoop Image Processing Interface) is an image processing library designed to be used with the Apache Hadoop MapReduce parallel programming framework. HIPI facilitates efficient and high-throughput image processing with MapReduce style parallel programs typically executed on a cluster. In the present work the HIPI library was modified, giving additional functionalities to read and process the GeoTIFF format (format provided by USGS -United States Geological Survey- for Landsat satellite images).
State of the Art There are different techniques used for image classification in semantic taxonomy categories such as vegetation, water, etc. (Codella et al., 2011), however these methods don’t consider scalability as part of its solutions, Noel C. F. Codella et. al.. In Wanfeng Zhang, et. al. (Zhang et al., 2013) An infrastructure for massive processing of satellite images in a multi-dataCenter environment, consisting of a DataCenter, where Access Security, Information Service and strategy Scheduling for data management us introduced. It’s important to consider HIPI (Sweeney et al., 2011) as a state of the art, extensible library for image processing and computer vision applications, which helps to avoid the problem of small files and achieving improvements in memory and response time.
Proposal That is why this paper is a modification of HIPI, to extends its functionality to work with TIFF images or GeoTIFF type. To achieve this, we proceeded as follows: • It was decided to use the Tiff format from satellite images obtained from the USGS since this format unlike others has no compression or data loss (Adobe, 1992). • JAI API was chosen to read and write the chosen format, JAI has more codecs and features available that can be useful for reading multiple formats. • Classes needed to upload, encode and decode
images of Tiff type were modified.
Based on the tests, the possibility of using HIPI for processing multispectral and hyperspectral images was analyzed. For such images, operations such as PCA are important to process all spectral bands, it is proposed to keep them in the same zip file, or as different images belonging to a tiff format, and then decode and interpret as a conventional multiband image. 4
Experiments The experiments were performed on a Local Heterogeneous Cluster depicted in Table 1 where the characteristics of slaves and master are shown. We used satellite images from LandSat 7, only considering the first 4 bands so we compressed a satellite image in .zip then we used the .zip up to 0.5GB, 1GB, 5GB and 10GB. The algorithm was tested about the average of channels which is explained in the official website of HIPI. In each task map, we iterated over each read of band of satellite image as FloatImage, added each value of pixel depending of channel then divided for number of pixels (width x height) and returned the key of the satellite image and array of data calculated. In each reduce task, we only calculated the average of the average of channels from each satellite image.
Node master slave 1 slave 2 slave 3 characteristics Core i7, RAM 8GB, Disk 100GB, S.O Ubuntu 64 bits Core i7, RAM 8GB, Disk 100GB, S.O Ubuntu 64 bits Core 2 Duo, RAM 4GB, Disk 100GB, S.O Ubuntu 64 bits Core 2 Duo, RAM 4GB, Disk 100GB, S.O Ubuntu 64 bits
In Table 2 shows in axis x the amount of data in GBs and in y axis the execution time. The Hadoop configuration was 1 replication of data, chunks of 32MB, 64MB and 128MB, 4096MB in memory for task reduce and map. The java virtual machine for task reduce and map was configured with 4096MB at the most. 5 Based on the review conducted and theoretical experimental tests HIPI modified version of the article concludes as follows: It is possible to perform various image processing operations, such as fil4,500 4,000 ]S3,500 [ e 3,000 m ti 2,500 n tio 2,000 cu 1,500 e xE1,000 350000 50 1 2 3 4 5 6 7 8 9 10 11 12
Amount of data [GB] Table 2: Execution time vs Amount of data spatial ters, variance, clustering or dimensionality reduction by using the MapReduce algorithm and also while the information in compressed format occupies less space, this does not necessarily mean faster times when processing, since the matrix calculations are done on the same decompression.
References 0xdata:H2O. 2015. H2o@ONLINE. Website. 0xdata:H2O, In: http://0xdata.com/ product/,accessed(may 2015).
Apache:Spark. 2015. H2o@ONLINE. Website. Spark, Lightning-fast cluster computing , In : https://spark.apache.org, accessed(accessed(2015-05-20)).
Noel C.F. Codella, Gang Hua, Apostol Natsev, and
John R Smith. 2011. Towards large scale land-cover recognition of satellite images. In Information, Communications and Signal Processing (ICICS) 2011, 8th International Conference on, pages 1–5.
IEEE.
Konstantin Shvachko, Hairong Kuang, Sanjay Radia, and Robert Chansler. 2010. The hadoop distributed file system. In Mass Storage Systems and Technologies (MSST), 2010 IEEE 26th Symposium on, pages 1–10. IEEE.
Chris Sweeney, Liu Liu, Sean Arietta, and Jason
Lawrence. 2011. Hipi: a hadoop image processing interface for image based mapreduce tasks.
Chris,University of Virginia.
Wanfeng Zhang, Lizhe Wang, Dingsheng Liu, Weijing
Song, Yan Ma, Peng Liu, and Dan Chen. 2013. Towards building a multi-datacenter infrastructure for massive remote sensing image processing. Concurrency and Computation: Practice and Experience, 25(12):1798–1812.
Multi-agent system for usability improvement of a university administrative system.
Jorge Leoncio Guerra Guerra
Universidad Nacional Mayor de San Marcos Facultad de Ingeniería de Sistemas e Informática Laboratorio de Robótica e Internet de las Cosas jguerrag@unmsm.edu.pe
Félix Armando Fermín Pérez
Universidad Nacional Mayor de San Marcos Facultad de Ingeniería de Sistemas e Informática Laboratorio de Robótica e Internet de las Cosas
fferminp@unmsm.edu.pe
Abstract The implementation of multi-agent systems is one of the specific ways to integrate heterogeneous information systems, by creating a software agent that performs specific tasks within a field of known action, as in the case of the travelling salesman problem or an agent searching clinical data of a patient. In this paper, the development of a mobile agent is proposed for the implementation of university administrative services across heterogeneous servers, interacting in turn with other intelligent agents in an heterogeneous communication environment. This will generate a single-access interface, which will serve to improve the usability of the system when the client makes an administrative request regardless of the server on which the requested procedure is processed.
Keywords: Integration, usability, agents, multiagent systems, heterogeneous services. 1. Introduction The portal of the University Inca Garcilaso de la Vega (http://www.uigv.edu.pe) is a web application that offers different services available in each of the offices of the institution, which is accessed through hyperlinks from Faculties, Research Units, Distance Education, etc.
An integrated system by software agents (Shiao, 2004); in which, with a single interface, a user can access different services without having to go through different pages to finish their request; will make possible to link these services regardless of the Faculty where they are and make the process requested by the user in a transparent manner, so that the usability of the system is also improved. 2. Proposed implementation The structure of the proposed multi-agent system is shown in Figure 1.
Figure 1 – Proposed multi-agent system Using the GAIA methodology (Zambonelli et al, 2003) for modeling a system of agents, agents that will be built on the prototype are defined: a.- Client Agent: captures requirements or service requests from users; interacts with the mobile agent to process the request and sends the response to the client system for viewing or joining the client process. b.- Server Agent: interacts with the server of the corresponding office to transfer the request from the mobile agent, processes the requested service, and sends the response to the mobile agent. c.- AGMER Agent: mobile agent moving through JADE middleware to interact with clients and servers, has the ability to divide received requests into subtasks performed by different servers.
One advantage of using GAIA is that it can be combined with diagrams from UML as described in Kang et al (2004), for this reason UML diagrams and GAIA models (Wooldridge et al is used., 2000) will be used for its construction.
Once developed, the model implemented as shown in Figure 2.
will
be
In the development process is suggested the use of Eclipse Luna, JDK 8.0_41 and JADE 4.3.2 for creating agents. In the prototype, it was initially considered the access to these basic services for a university student from the case study: a. Specialized Library System with Struts 2, running on Google App Engine cloud, a PAAS solution. b. PRODI (English Learning Program) registration system with Node.JS, OpenShift, a PAAS solution.
All modules will communicate via web services, with interface agents using AXIS 2. 3. Conclusions and future work The single-interface model for administrative procedure systems is suitable to improve the usability of a web portal to reduce the number of interactions.
A multi-agent system allows an heterogeneous implementation of components as well as asynchronous communication similar to a request of a university administrative procedure.
JADE is a suitable platform for developing agents due to its ease of use, and adaptability to development methodologies of multi-agent systems.
As future work it is proposed to develop indicators to measure implemented services, the frequency of use of the services, and others, that will be collected and stored in the cloud for future analysis using Big Data tools. 4. References Dan Shiao. 2004. Mobile Agent: New Model of Intelligent Distributed Computing. IBM China, October, 2004.
Franco Zambonelli, Nicholas Jennings, Michael Wooldridge. 2003. Developing Multiagent Systems: The Gaia Methodology. ACM Transactions on Software Engineering and Methodology. 12(3): 317–370.
Miao Kang, Lan Wang, Kenji Taguchi. 2004.
Modelling Mobile Agent Applications in UML2.0 Activity Diagrams, Proceedings of the 6th International Conference on Enterprise Information Systems. Porto. Portugal. Vol. IV: 519-522.
Michael Wooldridge, Nicholas Jennings, David Kinny. 2000. The Gaia Methodology for AgentOriented Analysis and Design. Journal of Autonomous Agents and Multi-Agent Systems, vol. 15. Kluwer Academic Publishers, Boston. The Netherlands.