=Paper=
{{Paper
|id=None
|storemode=property
|title=A Platform for Supporting Data Analytics on Twitter: Challenges and Objectives
|pdfUrl=https://ceur-ws.org/Vol-895/paper1.pdf
|volume=Vol-895
|dblpUrl=https://dblp.org/rec/conf/semweb/StavrakasP12
}}
==A Platform for Supporting Data Analytics on Twitter: Challenges and Objectives==
https://ceur-ws.org/Vol-895/paper1.pdf
A Platform for Supporting Data Analytics on Twitter:
Challenges and Objectives1
Yannis Stavrakas Vassilis Plachouras
IMIS / RC “ATHENA”
Athens, Greece
{yannis, vplachouras}@imis.athena-innovation.gr
Abstract. An increasing number of innovative applications use data from
online social networks. In many cases data analysis tasks, like opinion mining
processes, are applied on platforms such as Twitter, in order to discover what
people think about various issues. In our view, selecting the proper data set is
paramount for the analysis tasks to produce credible results. This direction,
however, has not yet received a lot of attention. In this paper we propose and
discuss in detail a platform for supporting processes such as opinion mining on
Twitter data, with emphasis on the selection of the proper data set. The key
point of our approach is the representation of term associations, user associa-
tions, and related attributes in a single model that also takes into account their
evolution through time. This model enables flexible queries that combine com-
plex conditions on time, terms, users, and their associations.
Keywords: Social networks, temporal evolution, query operators.
1 Introduction
The rapid growth of online social networks (OSNs), such as Facebook or Twitter,
with millions of users interacting and generating content continuously, has led to an
increasing number of innovative applications, which rely on processing data from
OSNs. One example is opinion mining from OSN data in order to identify the opinion
of a group of users about a topic. The selection of a sample of data to process for a
specific application and topic is a crucial issue in order to obtain meaningful results.
For example, the use of a very small sample of data may introduce biases in the out-
put and lead to incorrect inferences or misleading conclusions. The acquisition of data
from OSNs is typically performed through APIs, which support searching for key-
words or specific user accounts and relationships between users. As a result, it is not
straightforward to select data without having an extensive knowledge of related key-
words, influential users and user communities of interest, the discovery of which is a
manual and time-consuming process. Selecting the proper set of OSN data is im-
portant not only for opinion mining, but for data analytics in general.
1
This work is partly funded by the European Commission under ARCOMEM (ICT 270239).
In this paper, we propose a platform that makes it possible to manage data analysis
campaigns and select relevant data from OSNs, such as Twitter, based not only on
simple keyword search, but also on relationships between keywords and users, as well
as their temporal evolution. Although the platform can be used for any OSN having
relationships among users and user posts, we focus our description here on Twitter.
The pivotal point of the platform is the model and query language that allow the ex-
pression of complex conditions to be satisfied by the collected data. The platform
models both the user network and the generated messages in OSNs and it is designed
to support the processing of large volumes of data using a declarative description of
the steps to be performed. To motivate our approach, in what follows we use opinion
mining as a concrete case of data analysis, however the proposed platform can equally
support other analysis tasks. The platform has been inspired by work in the research
project ARCOMEM2, which employs online social networks to guide archivists in
selecting material for preservation.
2 Related Work
There has been a growing body of work using OSN data for various applications.
Cheong and Ray [5] provide a review of recent works using Twitter. However, there
are only few works that explore the development of models and query languages for
describing the processing of OSN data. Smith and Barash [4] have surveyed visualiza-
tion tools for social network data and stress the need for a language similar to SQL
but adapted to social networks. San Martín and Gutierrez [3] describe a data model
and query language for social networks based on RDF and SPARQL, but they do not
directly support different granularities of time. Mustafa et al. [2] use Datalog to model
OSNs and to apply data cleaning and extraction techniques using a declarative lan-
guage. Doytsher et al. [1] introduced a model and a query language that allow to que-
ry with different granularities for frequency, time and locations, connecting the social
network of users with a spatial network to identify places visited frequently by users.
However, they do not consider any text artifacts generated by users (e.g. comments
posted on blogs, reviews, tweets, etc.).
The platform we propose is different from the existing works in that we incorpo-
rate in our modeling the messages generated by users of OSNs and temporally evolv-
ing relationships between terms, in addition to relationships between users. Moreover,
we aim to facilitate the exploration of the context of data and enable users to detect
associations between keywords, users, or communities of users.
3 Approach and Objectives
We envisage a platform able to adapt to a wide spectrum of thematically disparate
opinion mining campaigns (or data analysis tasks in general), and provide all the in-
2
http://www.arcomem.eu/
frastructure and services necessary for easily collecting and managing the data. This
platform is depicted in Fig. 1, and comprises three layers.
OPINION MINING Opinion
DATA ANALYSIS
dynamics
Diversification
Sentiment
analysis
Target
View on term, temporal, metadata, and definition
community criteria
Integrated querying
INTEGRATED
MODELING
Temporal modeling
Graph
Entity recognition Term association
Database
Community modeling
Campaign manager
CAMPAIGN CRAWLING
Campaign Tweet Community
Repository crawler crawler
Campaign
configuration
Tweets Community
User and tweet data
Twitter metadata
Fig. 1: Platform architecture
The first layer is the Campaign Crawling layer. The objective of this layer is to al-
low for the definition and management of campaigns, and to collect all the relevant
“raw” data. A campaign is defined by a set of filters and a timespan. For each cam-
paign, the platform monitors the Twitter stream for the duration specified by the cam-
paign’s timespan, and incrementally stores in the Campaign Repository all the data
that match the campaign filters. Stored data fall into three categories: tweets, metada-
ta, and community data. Community data describe the relationships among Twitter
users, while metadata may refer either to tweets (timestamp, location, device, lan-
guage, etc.) or to users (place, total activity, account creation date, etc.). Selecting the
correct tweets for a campaign is of paramount importance; therefore, for converging
to the proper filters the process of campaign configuration follows an iterative ap-
proach: a preliminary analysis on an initial small number of tweets gives an overview
of the expected results, indicates the most frequent terms, and highlights the most
influential users, in order to verify that the campaign is configured correctly. If this is
not the case, an adjustment step takes place by modifying the terms that tweets are
expected to contain, the user accounts deemed most relevant to the topic of the cam-
paign, or by removing tweets from user accounts that have been identified to be ro-
bots. Those modifications are based on suggestions made by the platform according to
the preliminary analysis of tweets.
The second layer is the Integrated Modeling layer. The objective of this layer is to
support a complex and flexible querying mechanism on the campaign data, allowing
the definition of campaign views. The significance of campaign views will become
apparent in a motivating example that will follow. A prerequisite for querying is the
modeling of the campaign “raw” data. The model should encompass three dimen-
sions. First, represent associations between interesting terms contained in tweets.
Such terms can be hashtags, or the output of an entity recognition process. Associa-
tions between terms can be partly (but not solely) based on co-occurrence of the terms
in tweets. Second, represent associations between users of varying influence forming
communities that discuss distinct topics. Finally, it should capture the evolution
across time of the aforementioned associations, term attributes, and user attributes.
This temporal, often overlooked, dimension is of paramount importance in our ap-
proach since it enables the expression of time-aware conditions in queries. The result-
ing model integrates all three dimensions above in a unified way as a graph. A suita-
ble query language is then used to select the tweets that satisfy conditions involving
content, user, and temporal constraints.
The third layer is the Data Analysis layer. The query language is used to define a
“target” view on the campaign data. This “target” view corresponds to a set of tweets
that hopefully contain the answer to an opinion-mining question. This set is then fed
into a series of analysis processes, like diversification for ensuring the correct repre-
sentation of important attribute values, and sentiment analysis for determining the
attitude of users towards the question. Opinion dynamics is important for recognizing
trends across time.
Motivating example. A marketing specialist wants to learn what people say in Twit-
ter about Coca-Cola: (a) in cases when Pepsi-Cola is also mentioned, and (b) during
the Olympic Games. The first step is to define a campaign. He launches a preliminary
search with the keywords coca cola. An initial analysis of the first results reveals
other frequent terms and hashtags, and after reviewing them he decides to include the
following in the campaign definition: #cc, #cola and coke. Moreover, the platform
suggests groups of users whose tweets contain most often the relevant keywords. He
decides to include some of them in the campaign definition. Having set the campaign
filters, he sets the campaign timespan, and launches the campaign. The crawler down-
loads data periodically, which are incrementally stored in the Campaign Repository
and modeled in the Graph Database. The next important step for our marketing spe-
cialist is to define suitable “targets” that correspond to his initial questions. He does
that by using the query language of the platform for creating views on the campaign
data. An intuitive description of the queries for the two cases follows.
The first query returns the tweets that will hopefully reveal what people say about
Coca-Cola in cases when Pepsi is also mentioned, using the following steps: 1) find
the terms that are highly associated with Pepsi Cola, 2) return the tweets in which
Coca- and Pepsi-related terms are highly associated.
The second query return the tweets that will hopefully reveal what people say
about Coca-Cola during the Olympics in the following steps: 1) find the terms that are
highly associated with Olympic Games, 2) find the time periods during which those
terms are most often encountered, 3) find the groups of people that most often use
those terms during the specific time periods, 4) return the tweets of those people,
during the specified time periods, which mention the Coca-Cola search terms.
The final step is to conduct opinion-mining analysis on the two sets of tweets re-
turned by the queries. In our approach the emphasis is on selecting the most suitable
set of tweets, according to the question we want to answer. In our view, selecting the
proper set is paramount for opinion mining processes to produce credible results.
4 Technical and Research Challenges
There are several technical and research challenges that need to be addressed in the
implementation of the proposed platform, with respect to the acquisition of data, as
well as the modeling and querying of data.
Scalable Crawling. In the case that the platform handles multiple campaigns in paral-
lel, there is a need to optimize the access to the OSN APIs, through which data is
made available. Typically, APIs have restrictions in the number of requests performed
in a given time span. The implementation of the platform should aim to minimize the
number of API requests while fetching data for many campaigns in parallel. Hence,
an optimal crawling strategy is required to identify and exploit overlaps between
campaigns and to merge the corresponding API requests.
Temporal Modeling. A second challenge is the modeling of large-scale graphs,
where both nodes and edges have temporally evolving properties. Our approach is to
treat such graphs as directed multigraphs, with multiple timestamped edges between
two vertexes [6]. Given that the scale of the graphs can be very large both in terms of
the number of vertexes and edges, but also along the temporal dimension, it is neces-
sary to investigate efficient encoding and indexing for vertexes and edges, as well as
their attributes. We have collected a set of tweets over a period of 20 days using Twit-
ter’s streaming API. In this set of tweets, we have counted a total of 2,406,250 distinct
hashtags and 3,257,760 distinct pairs of co-occurring hashtags. If we aggregate co-
occurring pairs of hashtags per hour, then we count a total of 5,670,528 distinct pairs
of co-occurring hashtags. Note that the sample of tweets we have collected is only a
small fraction of the total number of tweets posted on Twitter. If we can access a larg-
er sample of tweets and consider not only hashtags but also plain terms, then the cor-
responding graphs will be substantially larger.
Advanced Querying. A third challenge is the definition of querying operators that
can be efficiently applied on temporally evolving graphs. Algorithms such as Pag-
eRank, which compute the importance of nodes in a graph, would have to be comput-
ed for each snapshot of the temporally evolving graph. While there are approaches
that can efficiently apply such algorithms on very large graphs [7], they do not con-
sider the temporal aspect of the graphs that is present in our setting. Overall, the im-
plementation of querying operators should exploit any redundancy or repetition in the
temporally evolving graphs to speed-up the calculations.
Data Analysis. The proposed platform allows users to define a body of tweets based
on complex conditions (“target definition” in Fig. 1), in addition to the manual selec-
tion of keywords or hashtags. Since the quality of any analysis process is affected by
the input data, a challenge that arises is the estimation of the bias of the results with
respect to the input data.
5 Conclusions and Future Work
In this paper we have proposed and described in detail a platform for supporting data
analytics tasks, such as opinion mining campaigns, on online social networks. The
main focus of our approach is not on a specific analysis task per se, but rather on the
proper selection of the data that constitute the input of the task. In our view, selecting
the proper data set is paramount for the analytics task to produce credible results. For
this reason we have directed our work as follows: (a) we are currently implementing a
campaign manager for crawling Twitter based on highly configurable and adaptive
campaign definitions, and (b) we have defined a preliminary model and query opera-
tors [6] for selecting tweets that satisfy complex conditions on the terms they contain,
their associations, and their evolution and temporal characteristics. Our next steps
include the specification and implementation of a query language that encompasses
the query operators mentioned above, and the integration of the implemented compo-
nents into the platform we described in this paper.
References
1. Y. Doytsher, B. Galon, and Y. Kanza. Querying geo-social data by bridging spatial net-
works and social networks. 2nd ACM SIGSPATIAL International Workshop on Location
Based Social Networks, LBSN '10, pages 39-46, New York, NY, USA, 2010.
2. W. E. Moustafa, G. Namata, A. Deshpande, and L. Getoor. Declarative analysis of noisy
information networks. 2011 IEEE 27th International Conference on Data Engineering
Workshops, ICDEW '11, pages 106-111, Washington, DC, USA, 2011.
3. M. San Martin and C. Gutierrez. Representing, querying and transforming social networks
with rdf/sparql. 6th European Semantic Web Conference on The Semantic Web: Research
and Applications, ESWC 2009 Heraklion, pages 293-307.
4. M. A. Smith and V. Barash. Social sql: Tools for exploring social databases. IEEE Data
Eng. Bull., 31(2):50-57, 2008.
5. M. Cheong and S. Ray. A Literature Review of Recent Microblogging Developments.
Technical Report TR-2011-263, Clayton School of Information Technology, Monash Uni-
versity, 2011.
6. V. Plachouras and Y. Stavrakas. Querying Term Associations and their Temporal Evolu-
tion in Social Data. 1st Intl VLDB Workshop on Online Social Systems (WOSS 2012), Is-
tanbul, August 2012.
7. U. Kang, D.H. Chau, C. Faloutsos: Mining large graphs: Algorithms, inference, and dis-
coveries. ICDE 2011: 243-254.