A Sentiment-Based Approach to Twitter User Recommendation

Abstract

Nowadays, the emerging popularity of Social Web raises new application areas for recommender systems. The aim of a social user recommendation is to suggest new friends having similar interests. In order to identify such interests, current recommender algorithms exploit social network information or the similarity of user-generated content. The rationale of this work is that users may share similar interests but have di erent opinions on them. As a result, considering the contribution of user sentiments, can yield bene ts in recommending possible friends to follow. In this paper we propose a user recommendation technique based on a novel weighting function, we named sentimentvolume-objectivity (SVO) function, which takes into account not only user interests, but also his sentiments. Such function allows us to build richer user pro les to employ in the recommendation process than other content-based approaches. Preliminary results based on a comparative analysis show the bene ts of the advanced approach in comparison with some state-of-the-art user recommender systems.

Categories and Subject Descriptors

H.3.3 [Information Search and Retrieval]: [Information Filtering]

1. INTRODUCTION

The growing popularity of social networks increases the availability of user sentiments, which has become a significant impact factor on buying decisions, brand reputations and public opinions. Furthermore, recommending pertinent news stories, documents, and users to follow, has long been a favourite domain for recommender systems research. Several new approaches harness real-time micro-blogging activity from services, such as Twitter1, as the basis for identifying user preferences and ltering relevant contents to speci c people. Recently, Twitter has become an interesting source of research activity as a result of the large amount of available user-generated data. In particular Twitter permits users to share a sentence - called tweet - to the followers, with a maximum length of 140 characters.

In this instance, the purpose of user recommendation is to identify relevant people to follow among millions of users that interact in the social network. Previous attempts include both content-based and graph-based approaches. The former focuses on metrics for measuring the topic similarity among Twitter users, the latter exploits the graph of relationships among users to infer correlations.

The main idea behind this work is that users may share similar interests but have di erent opinions about them. Therefore, we extend the content-based recommendation by means of the sentiments and opinions extracted from the user micro-posts in order to improve the accuracy of the suggestions. This leads us to de ne a novel weighting function in order to enrich content-based user pro les. 2.

RELATED WORK

In spite of the growing body of research on exploiting user-generated contents in recommendation engines, there are few attempts to consider sentiment included in microposts during the recommendation process. Singh et al. [ 17 ] introduce a hybrid recommender system that improves the results of collaborative ltering by incorporating a sentiment classi er in the movie recommendation scenario. Bank and Franke [ 4 ] try to better represent public product reviews on weblogs through di erent text mining techniques. Faridani [ 9 ] achieves the same goal by exploiting a multivariate regression approach. As far as we are aware, there are no attempts towards sentiment user recommendation in social networks.

User recommendation approaches that ignore user opinions have been proposed by Freyne et al. [ 10 ] and Chen et al. [ 8 ] exploring di erent recommendations strategies. Approaches for social recommendation that incorporate user opinions have been proposed in other domains, e.g., [ 5 ]. Guy et al. [ 12 ] propose a people recommendation engine within an enterprise social network site scenario. They aggregate several di erent sources to derive factors that might in uence the similarity measure. Twittomender [ 13 ] lets users nd pertinent pro les on Twitter exploiting di erent strategies, both content-based and collaborative ones. Arru et al. [ 3 ] propose a signal-based representation of user interests in order to draw similarities among people.

3. SENTIMENT ANALYSIS ALGORITHM

Sentiment analysis or opinion mining is formally de ned as the computational study of sentiments and opinions about an entity expressed in a text. According to Liu [ 15 ], the entity is classi ed into ve categories: product, person, brand, event, concept. Particularly, in this work we assume the concept as the sentiment analysis target entity. Sentiment analysis is a di cult task, hence - before the setup of the algorithm - some assumptions are needed. There are multiple granularity levels of sentiment analysis, as explained in [ 2 ]: feature-level, entity-level, sentence-level, document-level.

In this work we consider sentiment analysis at sentencelevel. Speci cally, in the Twitter domain we assume that a sentence matches the whole tweet. Moreover, we assume that each sentence contains only one opinion related to the entity.

The goal of our sentiment analysis system is to obtain an output value that represents how much positive, negative or neutral is the sentiment expressed in a tweet. For this reason, we implemented a Supervised Machine Learning algorithm based on a Nave Bayes classi er. With a view to training our algorithm, we needed a dataset with labeled tweets. However, due to the lack of a Twitter public dataset, we decided to follow an alternative approach. Instead of manually building a labeled dataset, Bhayani et al. [ 11 ] propose to employ a noisy dataset of positive, negative, and neutral tweets. The labels correspond to special sequences of characters in the tweets, such as positive or negative emoticons (e.g., :-D ;-( ), hashtags (e.g., #iloveit, #ihate) or keywords (e.g., good, sad). Even though these labels do not always correspond to the right sentiment expressed by the tweet, they allow us to collect a large amount of data for training. The Twitter API2 have been used to retrieve a set of tweets containing the aforementioned features. The nal training dataset counts 150000 tweets divided in 50000 tweets for each class. Because the experimental evaluation is conducted on events related to the 2013 Italian political elections, the TextCat language recognizer3 is employed to limit the set to Italian tweets. In order to increase the classi er precision and reduce the presence of noise, we performed a feature selection. In particular, the terms with low values of Salience are discarded. The Salience of a term t is de ned by Pak et al. [ 16 ] as follows: Salience(t) =

N likelihood that the term t belongs to the label class L. A zero value of Salience means that the term t appears uniformly in each dataset, thus it is a good candidate to be discarded. Finally, as for the Machine Learning algorithm, a Nave Bayes classi er is trained on the training data, where each tweet is represented as a feature vector made up of the following groups of features:

Bag-of-words: vectors of word unigram; Word polarities: using the LIWC4 content analysis dictionary, we extracted features for positive, negative, and neutral words. Individual word polarities are inverted if the word follows a negation; Negations: we add the "NEG " su x to each word following a negation pattern (e.g., "not perfect" becomes "perfect NEG"); Elongated words: we represent as a feature the presence of words with one character repeated more than two times, (e.g., "looove", "yesss"); Part-of-speech tags: they provide a rough measure of the tweet content.

4. SVO RECOMMENDATION APPROACH 4.1

User profiling

Several approaches to user recommendation are based on the de nition of a similarity measure between two users ui and uj. Given the user ui, the ranked list of suggested users corresponds to the set of users uj that maximize the aforementioned measure. Content-based approaches de ne this measure by analyzing the user tweets. The set T of tweets tweets(u) posted by the user u can be viewed as an extension of the bag-of-word model, where concepts are more semantically signi cant and less ambiguous than plain keywords. Instead of using complex semantic annotators, a concept is uniquely identi ed through hashtags contained in the tweet, namely, the metadata tags that are used in Twitter to indicate the context or the ow a tweet is associated with. Thus, we de ne the pro le p of the user u as the set of weighted concepts: p(u) = f(c; !(u; c))jc 2 Cug (2) where !(u; c) is the relevance of the concept c for the user u, and Cu is the set of concepts cited by the user u. The weighting function will be discussed in the following section.

The user pro le representation is generated by monitoring the user activity, that is, all the tweets included in the observation period. Afterwards, given two users ui and uj, and their pro les p(ui) and p(uj), the similarity function is de ned in terms of cosine similarity: = sim(ui; uj) = sim(p(ui); p(uj)) =

P

c2Cui [Cuj !(ui; c) !(uj; c) qPc2Cui !(ui; c)2 qPc2Cuj !(uj; c)2 (3) where Cui and Cuj are the concepts in the pro les of users ui and uj, respectively. 4liwc.net

The idea behind this work is that taking into account user attitudes towards his own interests can yield bene ts in recommending friends to follow. Speci cally, we consider (i) which is the sentiment expressed by the user for a given concept, (ii) how much he is interested in that concept, and (iii) how much he expresses objective comments on it.

In our model the rst contribution S(u; c), namely, the sentiment of the user u about a concept c, is obtained as follows:

S(u; c) = f

P os(u; c) N eg(u; c)

P os(u; c) + N eg(u; c) where P os(u; c) and N eg(u; c) are the sums of the positive and negative tweets written by the user u regarding the concept c, respectively. Such values are calculated by means of our proposed Machine Learning algorithm (see Section 3) that classi es the tweets as positive, negative or neutral. A low value of S(u; c) means that the user sentiments towards the concept c are negative, on the contrary a high value represents positive sentiments.

The f function is used to normalize the output value within the [0; 1] range: f (x) =

1 1 + k x where k = 10.

The second contribution is the volume V (u; c), that is, how much a user u wrote about a speci c concept c and is de ned as follows:

V (u; c) =

tweets(u; c)

PiN=1 tweets(u; ci) where tweets(u; c) is the number of tweets written by the user u about a speci c concept c, and N is the total number of concepts dealt with by u.

The third contribution is the objectivity O(u; c). With this term we denote how many tweets about a concept c do not contain sentiments or opinions and therefore may be objective. This may be important because objective tweets are typically news, so quite signi cant for the similarity of user pro les but less relevant for the sentiment analysis.

O(u; c) is de ned as follows:

O(u; c) =

N eutral(u; c) P os(u; c) + N eg(u; c) + N eutral(u; c) (7) where P os(u; c), N eg(u; c) and N eutral(u; c) are the sums of the positive, negative, neutral tweets written by the user u relative to the concept c, respectively.

Based on such contributions, we proposed a novel weighting function, we called sentiment-volume-objectivity (SVO) function, that takes into account all of them. It is de ned as follows:

SV O(u; c) =

S(u; c) + V (u; c) + O(u; c) (8) where , , and are three constants 2 [0; 1], such that + + = 1. The function SV O(u; c) 2 [0; 1] is the weighting function !(u; c) that appears in the Equations 2 and 3.

The experimental evaluations (Section 5) shows the computation of the values of the parameters , , and that maximize the performance of the recommender. (4) (5) (6) 5. 5.1

EXPERIMENTAL EVALUATION

Dataset

In order to evaluate the proposed model, we considered a case study rich of sentiments, such as the 2013 Italian political elections. Using the Twitter APIs we selected 31 hashtags for retrieving the Twitter streams about politician leaders and parties from Jan 25th to Feb 27th. Furthermore, because social networks are dynamic and fast-changing, we retrieved the hashtags that more often co-occur in the obtained tweets and added them to the initial hashtag set. This way, we took into account the trending topics that may be ignored in the initial query setup. The dataset counted 1085000 tweets, and over 25000 users that wrote almost one tweet. For the experimental evaluation we nally selected 1000 random users that (i) posted at least 50 tweets in the observed period, and (ii) had more than 15 friends and followers already stored into the dataset. The nal dataset for the evaluation counted 805956 tweets. 5.2

Evaluation

The goal of our user recommender system is to suggest to a user someone to follow, with similar interests and opinions. In order to compare di erent pro ling approaches and recommendation strategies, we need to understand when a user u1 is relevant for a user u2. In this work we suppose that u1 is relevant for u2 if a following relationship exists between them. This assumption has recently became a commonplace among social networks recommender systems [ 1, 14, 3 ] and is supported by the phenomenon of homophily, that is, the tendency of individuals with similar characteristics to associate with each other.

We performed a preliminary evaluation in order to assess the e ectiveness of the proposed approach. For the sake of brevity, in this paper we only report the results of a comparative analysis of our approach with two traditional approaches that do not consider sentiment: (i) cosine similarity in a Vector Space Model (VSM) where vectors are weighted hashtags, and (ii) the function S1 proposed by Hannonet al. [ 13 ]. We used di erent metrics to express the evaluation results. Success at Rank K (S@K) provides the mean probability that a relevant user is located in the top K positions of the list of suggested users. Mean Reciprocal Rank (MRR) indicates the average position of a user in the recommended list. Mean Average Precision at cut-o K (MAP@K) is the average of the precision value for each of the top-K recommended users. Figure 1 shows the obtained evaluation results. As can be seen, our approach outperforms the other ones according to each evaluation metric. These ndings con rm that sentiment is a valuable feature to be considered in order to improve the user recommender systems. As a marginal note, the absolute values of the achieved results are high due to the characteristics of the built dataset, where the relations among users are signi cantly dense. Finally, we also analyzed the user recommender performance in terms of variations of the three parameters , , and (see equation 8). In order to determine the best values of those parameters, we implemented a mini-batch gradient descent algorithm. The best results, according to aforementioned metrics, was achieved running the evaluation with = 0:3, = 0:6, and = 0:1. Based on the proposed model and the used dataset, these weights appear to highlight the contribution of the volume and the sentiment in comparison with the objectivity.

CONCLUSIONS

In this paper we have described a user recommender system for Twitter. Our work emphasizes the use of implicit sentiment analysis in order to improve the performance of the recommendation process. We have de ned a novel weighting function that takes into account sentiment, volume, and objectivity related to the user interests. This technique allowed us to build more complete user pro les than traditional content-based approaches. Preliminary results show the bene ts of our proposed model compared with some state-of-the-art methods.

As future work we are planning a deep sensitivity analysis to investigate whether social interactions, user preference and dataset characteristics shape parameters , , and . We will also include some improvements of the recommendation process taking into account other elements (e.g., named-entities, persons, products) and semantic representations of hashtags (e.g., [ 6 ][ 7 ]). A future study will also focus on the use of the implicit sentiment analysis within the collaborative ltering in social networks.

References

[1] F. Abel , Q. Gao , G.-J. Houben , and K. Tao . Analyzing temporal dynamics in twitter pro les for personalized recommendations in the social web . In Proceedings of ACM WebSci '11, 3rd International Conference on Web Science , Koblenz, Germany. ACM, June 2011 .

[2] A. Agarwal , B. Xie , I. Vovsha , O. Rambow , and R. Passonneau . Sentiment analysis of twitter data . In Proceedings of the Workshop on Languages in Social Media, LSM '11 , pages 30 { 38 , Stroudsburg , PA, USA, 2011 . Association for Computational Linguistics .

[3] G. Arru , D. Feltoni Gurini , F. Gasparetti , A. Micarelli , and G. Sansonetti . Signal-based user recommendation on twitter . Social Recommender Systems 2013 , 2013 .

[4] M. Bank and J. Franke . Social networks as data source for recommendation systems . In F. Buccafurri and G. Semeraro, editors, E-Commerce and Web Technologies , volume 61 of Lecture Notes in Business Information Processing , pages 49 { 60 . Springer Berlin Heidelberg, 2010 .

[5] C. Biancalana , F. Gasparetti , A. Micarelli , and G. Sansonetti. An approach to social recommendation for context-aware mobile services . ACM Trans. Intell. Syst. Technol. , 4 ( 1 ): 10 :1{ 10 : 31 , Feb . 2013 .

[6] C. Biancalana , F. Gasparetti , A. Micarelli , and G. Sansonetti . Social semantic query expansion . ACM Trans. Intell. Syst. Technol. , 4 ( 4 ), 2013 . Forthcoming issue.

[7] C. Biancalana and A. Micarelli . Social tagging in query expansion: A new way for personalized web search . In CSE (4) , pages 1060 { 1065 . IEEE Computer Society, 2009 .

[8] J. Chen , W. Geyer , C. Dugan , M. Muller , and I. Guy. Make new friends, but keep the old: recommending people on social networking sites . In Proceedings of the 27th International Conference on Human Factors in Computing Systems, CHI '09 , pages 201 { 210 , New York, NY, USA, 2009 . ACM.

[9] S. Faridani . Using canonical correlation analysis for generalized sentiment analysis, product recommendation and search . In Proceedings of the fth ACM Conference on Recommender systems, RecSys '11 , pages 355 { 358 , New York, NY, USA, 2011 . ACM.

[10] J. Freyne , M. Jacovi , I. Guy , and W. Geyer . Increasing engagement through early recommender intervention . In Proceedings of the third ACM Conference on Recommender Systems, RecSys '09 , pages 85 { 92 , New York, NY, USA, 2009 . ACM.

[11] A. Go , R. Bhayani , and L. Huang . Twitter sentiment classi cation using distant supervision . Processing , pages 1 { 6 , 2009 .

[12] I. Guy , I. Ronen , and E. Wilcox . Do you know?: recommending people to invite into your social network . In Proceedings of the 14th International Conference on Intelligent User Interfaces , IUI '09 , pages 77 { 86 , New York, NY, USA, 2009 . ACM.

[13] J. Hannon , K. McCarthy , and B. Smyth . Finding useful users on twitter: twittomender the followee recommender . In Proceedings of the 33rd European Conference on Advances in Information Retrieval, ECIR'11 , pages 784 { 787 , Berlin, Heidelberg, 2011 . Springer-Verlag.

[14] H. John , B. Mike , and S. Barry . Recommending twitter users to follow using content and collaborative ltering approaches . RecSys'10 : Proceedings of the 4th ACM Conference on Recommender Systems ,, 26 - 30 ( 10 ): 8 , 09 2010 .

[15] B. Liu . Sentiment analysis and subjectivity . Handbook of Natural Language Processing, , pages 627 { 666 , 2010 .

[16] A. Pak and P. Paroubek . Twitter as a corpus for sentiment analysis and opinion mining . volume 2010 , pages 1320 { 1326 , 2010 .

[17] V. K. Singh , M. Mukherjee , and G. K. Mehta . Combining collaborative ltering and sentiment classi cation for improved movie recommendations . In Proceedings of the 5th International Conference on Multi-Disciplinary Trends in Arti cial Intelligence , MIWAI'11 , pages 38 { 50 , Berlin, Heidelberg, 2011 . Springer-Verlag.