A Context-aware Recommender System for Hyper-local News - A Conceptual Framework Anh Nguyen Duc anhn@idi.ntnu.no 0 Hilde Gudvangen hilde@muml.community Department of Computer and Information Science , NTNU

Recommender systems (RSs) have been popular in variety of application domains due to the increased demand for filtering and sorting items and information. Today, there is a numerous approaches and algorithms of data filtering and recommendations. This works presents a conceptual framework for constructing a mobile RS in hyper-local news domain. The mobile RS is designed to deal with specific requirements of news readers, such as spatialtemporal relevance, recency, real-time update and validated news. The implementation of the RS in a distributed file system is also discussed.

 Recommeder system Collaborative Filtering Social filtering Hyper-local news real-time recommender system reinforcement learning Big Data 1 Introduction In Big Data era, information has increased at an unprecedented rate and the

information overload problem has become increasingly severe for online users.

 Nighty percent of all the data available today were created in the last two years [30]. In this context, RS plays an important role to bring meaningful and relevant

information to individuals and business organizations [ 1 ]. Starting from mid 1990s,

RSs became an independent research area with a large application domain, including

e-commerce, multimedia, work and productivity, news, education and tourism [ 1-3 ].

This work investigates the feasibility of applying a RS to a hyper-local news mobile

application.

The mission of a hyper-local news editor is to deliver relevant news to users as quick as possible, considering its location context. Hyper-local news is targeted at or consumed by people or entities that are located within a well-defined area, generally on the scale of a street, a neighborhood, a community or a city. Hyperlocal content must also be relevant in time. The higher the content scores on these dimensions the more relevant the content becomes to the individual and the less it becomes to the masses.

 Copyright held by the author(s). NOBIDS 2016 Mobile apps for hyper-local news are becoming popular in software startup scenes

around the world, such as BlogFeed1, Ripple2 and MittMedia3 to name a few.

 However, these startups are also facing with challenges of making sense out of the

large volume of data occurring in a real-time manner. We are particularly interested in

 RSs for mobile application. Besides the mentioned concerns, mobile RSs face a

challenge of making accurate recommendations using simple, yet appealing user interface [ 27 ]. Most of the mobile RSs heavily rely on locations of the users to recommend items to them, which is also essential in hyper-local news domain.

Alternatively, the recommendation is made based on not only item’s content but also user’s context variables, i.e. geographical location and time. This paper proposes a conceptual architecture of a mobile news RS applied for

hyper-local news with user-generated contents and social network data. Our solution adopts different recommendation techniques and considers mobile-specific factors, such as geographical location and temporal information. We also discussed the proposed RS in big data perspective.

 The paper is structured as follows: Section 2 presents background about recommendation approaches and challenges in recommending news. Section 3 describes requirements to a hyper-local news RS and Section 4 discuses its conceptual architecture. Finally, Section 5 concludes the paper with future perspectives. 2 Background 2.1. Recommendation techniques State-of-the-art recommendation approaches can be summarized as in Table 1.

Traditional recommendation approaches are classified into content based filtering, collaborative filtering and hybrid approaches [ 1,2 ]. Content-based filtering [ 6-8 ] utilizes several characteristics of a rated item to recommend a future item with similar characteristics. For example, a user selects a movie with a specific genre, IMDb score, and editor’s evaluation. A content-based RS will probably recommend a movie with the same genre, IMDb score and editor’s judgment for the user. Certain characteristics of an item, i.e. textual description, linked images or sound can be analyzed to find the similarity among items.

Collaborative filtering (CF) produces a recommendation using both user’s past

preference and also similar decisions made by other users [ 3-5 ]. The CF technique can be divided into user- based and item-based CF approaches [ 9 ]. In the user-based

CF approach, a user will receive recommendations of items liked by similar users. In

the item-based CF approach, a user will receive recommendations of items that are similar to those they were preferred in the past. Hybrid approaches combine multiple

 RS techniques to achieve a synergy between them. Several researchers have

attempted to combine CF and content-based approaches in order to smoothen their 1 http://blockfeed.com 2 https://ripple.co 3 https://www.mittmedia.se disadvantages and to gain better performance [ 10-13 ].

Recent advancement in RS considers additional characteristics of user reading [14

18]. Social-based RSs utilize social interactions among users, which are available in

 Internet, to improve the effectiveness of traditional recommendation mechanisms. The

social interactions include online friending, making social comments, social tags, etc.

 Other types of social relations are also used for recommendation generation, i.e. social

bookmarks, physical context, social tags, and “co-authorship” relations [ 14-18 ].

Particularly, trust based system puts a weight on the opinions of an user which is a

friend or a person that the user can trust [ 14, 15 ].

Context-aware RSs utilize information such as time, geometrical information, or the company of other people (friends, families or colleagues for example), for some applications in which it is not sufficient to consider only users and items, such as recommending a vacation package, or personalized content on a website [ 2 ]. For example, using a temporal context, a travel recommender system might make a very different vacation recommendation in winter compared to summer [ 19 ]. The contextual information about users in technology enhanced learning environments is also incorporated into the recommendation process [ 20 ].

Group recommender system (GRS) applies for a group of users as a unit of

analysis instead of an individual. While many RSs are focused on making recommendations to a single user, many daily activities such as watching a movie or going to a restaurant involve a group of users, in which case recommendations must take into account the tastes and preferences of all the users in the group [ 1 ].

2.2. The challenges of recommeding a hyper-local news

News RS do not escape the common challenges with general RSs [ 21,26 ]. Özgöbek et al. discussed several issues in a news recommender system [ 21 ], as shown in Table 1. We added three news challenges that are specifically important for hyperlocal news domain, namely context awareness, social concern and real-time update (notated by * in Table 2).

• •

Context awareness: in hyper-local news domain, readers would like to read news that are relevant to his location, i.e. what interesting things happen in a nearby street corner. When travelling, the information about traffic jam does only matter if it is on user’s way.

Social relation concern: readers would be more interested to read news from their friends or the ones they follow. This would require the integration of personal data from social networks. •

Real-time update: most of the traditional RSs have two phases, offline model construction and on-demand recommendation phase, where the model is fed with new data. The model can be updated at regular time intervals, e.g., hours or days, cannot meet the real-time demands [ 22, 23 ].

3. Requirements to a hyper-local news RS

Muml AS is a software startup located in Trondheim. The company develops a hyper-local news service to provide users with relevant and validated news in the realtime manner. Users who install the mobile app can be notified news that are of their interest happening nearby. The company has gone through initial startup phases by (1) refining the product ideas, (2) validating the market demand at different scales, (3) and building up the first Minimum viable product (MVP), a technically demonstrable version of the product. In a large city like Singapore, Kuala Lumpur and Hanoi, a vast various types of news, including shopping, events, traffic, neighborhood, restaurant and services were available for publish from every street corner. A preliminary market study on a local community4 recorded thousands of post per hour. Big data issue is soon a challenge for scaling up the product.

Muml plans to embedd an advanced feature in the second MVP, so-called SmartGuide. The feature aims to provide intelligence for the service, by

recommending news to an user in a real-time manner. Each piece of news is collected with the following attributes: (1) a photo/ a short video, (2) description, (3) category, (4) channel, (5) hashtag, (6) location, (7) time created and (8) user created. User’s usage log is stored in a device and updated to a server frequently. The usage log includes: (1) read news list, (2) likes list and (3) comments.

Requirements for SmartGuide system has been set as below: • The system should collect user’s historical usage data from the user’s device. • Usage data should be associated with context information, i.e. user’s location, timestamp. • The system should, based on the usage data, recommend relevant news to the user. • The relevance should consider factors: time, location, user’s preference and user’s friend’s preference. • The system should provide a real-time update. • The system should deliver the recommended news queries, i.e., by users requests (pull-based).

 4. Conceptual architecture Muml is implemented as a standard mobile application with a client-sever architecture via REST. The frontend part is thin, presenting news to users in both a map view and a list view. The frontend has a cache to support store offline data and usage log. The backend implements all main logic functions, including the

SmartGuide feature. The news is recommended to a user based on his reading history as well as other user’s preference. Searching for relevant news in a greedy manner often impact on a long-term performance. We adopt a reinforcement learning approach to balance the tradeoff between exploration and exploitation. The main element of SmartGuide is shown as in Figure 1. 4 beat.vn Content analyzer component analyzes the description part of a news, mapping the identified textual elements into predefined topics, as described in a previous research [ 28 ]. This information combines with other attributes (as in Section III) to provide a news profile.

Profile learner component updates a news profile database and a user profile database when exploring new data entering to the system. The user profile is initially empty. Based on a reinforcement algorithm [ 31 ], when user select a news to read, the user profile is updated, either by adding new preferences or by updating reward/punishment values associated to existing elements of the profile. The specific algorithm in charge of managing these punishments/rewards is Q-learning [ 32 ]. The core part is a storage of state-action pairs and we can learn from the changing of value state Q(s,a) between state- action pair to another state-action pair. An abstract news content can be viewed as states s of the system and moving can be viewed as action a of state:

Q(s,a )= Q(s,a )+ α[rt+1 + γQ(st+1,at+1) − Q(st ,at )] Change detection component calculates two types of preference for each user, shortterm and long-term preference. The change is detected when there is a derivation of short-term interest’s value [ 29 ]. Particularly, more weight is given to a news category that has drawn recent user interest.

Spatial-temporal filtering component limits the exploration domain by excluding news that are out of the interest scope, i.e. news that are outside 5 km radius and beyond 24 hours. The current configuration is predetermined for a metropolitan context. It is can also configured for other context setting.

Context factors weighting component incorporates the influences of contextual factors, so that the trendy news will be weighted more, the news read by user’s friends will be weighted more and the news are of user’s preference will be weighted more. The framework is designed to solve the recency issue by giving more weights to the more recent news. We also consider diversity of news by adjusting weights given alongside trends, categories and social relations.

Perspective on Big data in our architecture is the deployment of data storage and query in distributed file system platforms. i.e. Hadoop and Spark. Three elements will be parallel processed in the distributed file systems: (1) Content analyzer, (2) Profile learner and (3) Context factor weighting. Data processing will eventually end in storing two tables: New_Similarity (News_ID, Similar_News, Similarity_Score) and User_News_Base (User_ID, News_ID, Recommendation_Score). The implementation of the architecture is split into an offline process and online process. The data preprocessing and analysis of news usage log, user profile, news database, and social relation is done in batch in a daily/ weekly basis. The online process composes querying new similarity and user news base tables. Online data processing includes the handling of the weekly results and also the most recent updated user read log.

 6. Discussion and Conclusion

[3] G. Adomavicius, A. Tuzhilin, Toward the next generation of recommender systems: a survey of the state-of-the-art and possible extensions, IEEE Transactions on Knowledge and Data Engineering vol. 17 (6), 2005, pp. 734–749

 [1] L. Boratto , S. Carta , State-of-the-art in group recommendation and new approaches for automatic identification of groups , in: G. Alessandro Soro, Vargiu E.P.G . (Eds.), Information Retrieval and Mining in Distributed Environments , Studies in Computational Intelligence , vol. 324 , Springer, Berlin, Heidelberg, 2011 , pp. 1 - 20 . [2] J. Lu , D. Wu , M. Mao , W. Wang , G. Zhang, Recommender system application developments: A survey, Decision Support Systems , vol 74 , 2015 , pp. 12 - 32 [4] J.L. Herlocker , J.A. Konstan , A.L. Borchers , J.T. Riedl , An algorithmic framework for performing collaborative filtering , 22nd Annual International Conference on Research and Development in Information Retrieval , 1999 , pp. 230 - 237 . [5] J. L. Herlocker , J.A. Konstan , J.T. Riedl , L.G. Terveen , Evaluating collaborative filtering recommender systems , ACM Transactions on Information Systems , vol. 22 ( 1 ) , 2004 , pp. 5 - 53 . [6] J. Bobadilla , F. Ortega , A. Hernando , A. Gutiérrez , Recommender systems survey , Knowledge-based Systems , vol( 46 ), 2013 , pp. 109 - 132 [7] R. J. Mooney , L. Roy, Content-based book recommendation using learning for text categorization , Workshop Recom. Sys.: Algo. and Evaluation , 1999 [8] R. Meteren , M. Someren , Using content-based filtering for recommendation , in: Proceedings of ECML 2000 Workshop: Maching Learning in Information Age , 2000 , pp. 47 - 56 [9] B. Sarwar , G. Karypis, J. Konstan , J. Riedl , Item-based collaborative filtering recom- mendation algorithms , Proceedings of the 10th International Conference on World Wide Web, ACM 2001 , pp. 285 - 295 . [10] R. Burke , Hybrid web recommender systems , in: P. Brusilovsky , A. Kobsa , W. Nejdl (Eds.), The Adaptive Web , Springer-Verlag, Berlin Heidelberg 2007, pp. 377 - 408 . [11] B. Mobasher , X. Jin , Y. Zhou , Semantically enhanced collaborative filtering on the web , in: B. Berendt , A. Hotho , D. Mladenič , M. Someren , M. Spiliopoulou , G. Stumme (Eds.), Web Mining: From Web to Semantic Web, Springer, Berlin Heidelberg 2004, pp. 57 - 76 . [12] B. Smyth , P. Cotter , A personalised TV listings service for the digital TV age , Knowledge-Based Systems 13 , 2000 , pp. 53 - 59 . [13] D. Billsus , M. Pazzani , User modeling for adaptive news access, User Modeling and User-Adapted Interaction , 2000 , pp. 147 - 180 . [14] D. Ben-Shimon , A. Tsikinovsky , L. Rokach , A. Meisles , G. Shani , L. Naamani, Recommender system from personal social networks , Advances in Intelligent Web, Mastering , Springer, 2007 . 47 - 55 . [15] C.-S. Hwang , Y.-P. Chen , Using trust in collaborative filtering recommendation , New Trends in Applied Artificial Intelligence, LNCS vol 4570 , 2007 , pp. 1052 - 1060 [16] K. Shiratsuchi , S. Yoshii , M. Furukawa , Finding unknown interests utilizing the wisdom of crowds in a social bookmark service , 2006 IEEE/WIC/ACM International Conference on Web Intelligence and Intelligent Agent Technology , 2006 , pp. 421 - 424 . [17] H. Ma , T.C. Zhou , M.R. Lyu , I. King , Improving recommender systems by incorporat- ing social contextual information , ACM Transactions on Information Systems (TOIS) , vol. 29 , 2011 , p. 9 . [18] S.-Y. Hwang , C.-P. Wei , Y.-F. Liao , Coauthorship networks and academic literature recommendation , Electronic Commerce Research and Applications , vol 9 , 2011 , pp. 323 - 334 . [19] S. Stabb , H. Werther , F. Ricci , A. Zipf , U. Gretzel , D.R. Fesenmaier , C. Paris, C. Knoblock, Intelligent systems for tourism , IEEE Intelligent Systems , vol. 17 , 2002 , pp. 53 - 66 . [20] K. Verbert , N. Manouselis , X. Ochoa , M. Wolpers , H. Drachsler , I. Bosnic , E. Duval, Context-aware recommender systems for learning: a survey and future challenges , IEEE Transactions on Learning Technologies , vol 5 ., 2012 , pp. 318 - 335 [21] O. Özgöbek , J. A. Gulla , R.C. Erdur , A survey on Challenges and Methods on News Recommendation , WEBIST , 2014 , pp. 278 - 285 [22] B. Chandramouli , J. J. Levandoski , A. Eldawy , M. F. Mokbel , StreamRec: A Real-Time Recommender System , 11th International Conference on Management of data , 2011 , pp. 1243 - 1246 [23] Y. Huang B. Cui W. Zhang J. Jiang Y. Xu , TencentRec: Real-time Stream Recommendation in Practice , SIGMOD'15 , pp. 227 - 238 [24] T. Belluf , L. Xavier , R. Giglio , Case study on the business value impact of personalized recommendations on a large online retailer , RecSys 12 , 2012 , pp. 277 - 280 [25] C. A. Gomez-Uribe , N. Hunt , The Netflix Recommender System: Algorithms , Business Value, and Innovation, Journal ACM Transactions on Management Information Systems , vol 6 ( 4 ), 2016 [26] J. A. Gulla1 , A. D. Fidjestøl1 , X. Su , Castejon., Implicit User Profiling in News Recommender Systems, 10th International Conference on Web Information Systems and Technologies , 2014 , pp. 185 - 192 [27] H. van der Heijden , G. Kotsis, R. Kronsteiner , Mobile recommendation systems for decision making “on the go‟ . In Proceedings of the International Conference on Mobile Business (ICMB'05) , 2005 [28] J. E. Ingvaldsen and J. A. Gulla , Taming News Streams with Linked Data , 2015 . [29] H.J. Lee , S. J. Park , MONERS: A news recommender for the mobile web , Expert Systems with Applications , vol. 32 ( 1 ), 2007 , pp. 143 - 150 [30] C. Dobrea , and F. Xhafa , Intelligent services for Big Data science , Future Generation Computer Systems , vol. 37 , 2014 , pp. 267 - 281 . [31] Y. Seo , B. Zhang, A reinforcement learning agent for personalized information filtering , IUI 2000 , pp. 248 - 251 [32] P. Rojanavasu , P. Srinil , O. Pinngern , New Recommendation System Using Reinforcement Learning , International conference on eBusiness, Bangkok, 2005