Recommender systems in the era of sentiment analysis and social media

Abstract

Recommender Systems (RSs) are software tools and techniques providing suggestions for items to be of use to a user . The suggestions is regarding various decision-making processes, such as what items to buy, what hotel to choose, or what user connect with into social networks. The information can be acquired explicitly (typically by collecting users’ ratings) or implicitly by monitoring users’ behavior, such as songs heard, applications downloaded, users’ social timeline, and web sites visited). RS may use demographic features of users (like age, nationality, gender), psychological features, and social information, like followers, followed, tweets and posts, that are commonly used in Web 2.0. It is also growing the use of information from Internet of things (e.g., GPS locations, RFID, real-time health signals). As a matter of fact, RS implementation in the Internet has recently increased, which has facilitated its use in di↵erent areas. The most common research papers are focused on movie recommendation studies, however a great volume of literature for RS is centered on di↵erent topics, such as music, television, books, documents, scientific documents, e-learning, e-commerce in markets and social media, among others. Social media have become very popular in recent years because of the increasing proliferation of Internet enabled devices such as personal computers, smartphones, and mobile devices. This is evidenced by the promising popularity of many online social networks such as Twitter, Facebook, and LinkedIn. Such social networks have lead to a tremendous explosion of network-centric data in a wide variety of fields. In this scenario, with the expanding demand of RSs on social networks, detecting sentiments and opinions from the Web is becoming an increasingly widespread and important form of data interpretation. Sentiment Analysis (SA) or Opinion Mining aims to understand subjective information, such as opinions, points of views and feelings expressed by users in the content they generate. SA algorithms identify how positive, negative or neutral is the produced content regarding a specific entity, that is, a product, person, organization, event and topic. As a matter of fact, Sentiment Analysis in the field of social media permits companies, marketers, organizations, or individuals, to understand their business online reputation, identify public opinions regarding products and services of themselves or their competitors, and gain insights about possible emerging trends and changes in market opinions, or identify crisis. During the Ph.D. I was trying to combine these three research area. First of all, I’ve studied all the most used techniques for Recommender Systems with a focus to people-topeople recommendation, especially for the social network recommendation purpose. I have also focused on the study of data mining algorithms to extract data from social networks and to store it into database for future analysis. I deeply researched how user relationships are formed in social networks and which graph or clustering algorithms are more e↵ective to use in combination with a Recommender System. Furthermore I’ve researched methods and challenges regarding Sentiment Analysis. I devised a simple proprietary algorithm of sentiment analysis tailored for Twitter, to make use in my own researches. The main research goal was to understand if Sentiment Analysis on social networks may help the recommendation process, and in particular improving the precision and other non-accuracy measures (novelty and diversity) of the recommender. The following list provide a brief description of the contributions of my Ph.D. with respect to the research field discussed above. Sentiment-based User Recommender on Twitter. In this work is proposed a new weighting function that takes into consideration Sentiment Analysis of posts, with the aim to improve the recommendation task. The rationale behind this work is that users in social networks may share similar interests but might have di↵erent opinions on them. As a result, considering the contribution of user sentiments can yield benefits in recommending possible friends to follow. Firstly, we devised a proprietary algorithm of sentiment analysis, that is specific for Twitter analysis. Secondly, we propose a user recommendation technique based on a di↵erent weighting function, we named sentiment-volume-objectivity (SVO) function, which takes into account not only user interests, but also sentiments toward them. Such function allows us to build richer user profiles to employ in the recommendation process than other content-based approaches. The main research question we advance is: can the consideration of this novel sentiment-based function yield benefits to the user RS? Exploiting Signals and Temporal Dynamics for a People-to-People RS. In this part is introduced a novel framework with a new user model, called bag-of-signals, that represents how user interests vary over time for creating more comprehensive user profiles. The basic idea underlying such approach is to represent each user interest as a signal. In order to analyze such signals we make use of the wavelet transform, a signal processing technique that captures the frequency content of any signal, together with their precise location of occurrence in the time domain. After evaluating the performance of this techniques we consider another signal dimension that represents the sentiment of a user toward a specific topic. The Sentiment Analysis model is build as the previous section Sentiment-based User Recommender on Twitter. The research questions we pose are: (i) can the consideration of temporal patterns of changing users interests really impact the characteristics and quality of user recommender? (ii) Can Sentiment Analysis yield some benefits to the proposed temporal-based RS? Leveraging Community Detection Techniques for User RS. From the evaluation results of the previous works, Sentiment Analysis has preliminary proved its benefits for a people RS. In this work we want therefore to build a more complex user Recommender System that can exploit the potentials of SA and social networks, considering also how topic evolve and change in user comments. To reach this goal, we propose a new approach for realizing user recommenders, named SCORES (Sentiment COmmunities REcommender System.) This algorithm relies on the identification of sentiment communities in which, for each topic cited by the user, we consider not only the relative sentiment, but also the SVO of contents generated by him. The graph is built by considering each topic discussed by the users as a vertex and the edges are generated by considering the Tanimoto similarity between users. Clustering based on the modularity optimization allows us to detect the latent communities. The recommendation process occurs by suggesting to the target user the most similar K users based on several tie strength measures. The research questions we set are: (i) which is the best graph techniques that enhance the contribution of the Sentiment Analysis? (ii) Can sentiment improve the final recommendation precision? (iii) Are there di↵erences depending on the category of topics dealt with by the user? Matrix Factorization Recommender System. To address scalability issues and temporal dynamics we propose a novel recommendation engine, that still relies on the identification of semantic attitudes, that is, sentiment, volume, and objectivity extracted from user-generated content. In order to do this at large-scale on traditional social networks, we devise a three-dimensional matrix factorization, one for each attitude. Potential temporal alteration of users’ attitudes are also taken into consideration in the factorization model. This work also represents one of the first attempt to combine sentiment in a matrix factorization recommender systems. Research questions that we want to answer are (i) does content published by users and, in particular, the inferred attitudes, allows for a better identification of potential relationships between users? How does temporal analysis of these attitudes impact the recommendation? Furthermore the scientific contributions coming from this section also include a comparative experimental results of a set of di↵erent evaluation metrics, including a range of non-accuracy measures, such as diversity and novelty and an extensive evaluation of the proposed algorithm on real world datasets. A Sentiment-based Youtube Video Recommender. To understand whether Sentiment Analysis enriches the recommendation process also in others social networks, we explore similar techniques for video recommendation on Youtube. Youtube social network is a specific video sharing network, where comments left by the viewers often provide valuable information to describe sentiments, opinions and tastes of the users. For this reason, we propose a novel re-ranking approach that takes into consideration that information in order to provide better recommendations of related videos. The research question we pose in this Chapter is therefore: how much Sentiment Analysis can enrich the recommendation process on Youtube? A preliminary evaluation highlights an increase of the recommender precision compared with a state-of-the-arts approach.