A Framework for Opinion Change Mining in Social Networks: Tracking Opinion change on Twitter

Un cadre de fouille de données pour la détection de changement d'opinion dans les réseaux sociaux: le cas des changements d'opinion sur Twitter

Jury de thèse:

• Prof. Giovanna DI MARZO SERUGENDO, directrice de thèse, Université de Genève 
• Prof. Jean-Henry MORIN, président du jury, Université de Genève 
• Prof. Thang LE DINH, Université du Québec à Trois-Rivières 
• M. Mehdi SNENE, Université de Genève

Résumé:

The increased prominence of social media in society has prompted studies in many research fields, such as social network analysis (SNA), natural language processing (NLP), data mining, etc. Issues worthy of study include cascades in the network (i.e. shapes of re-tweets), information diffusion, influencer detection, recommendation systems or opinion mining. These studies provide coarse-grained tracking of opinions across an overall network. As of yet, there are no specific studies focusing on fine-grained opinion changes tracking, e.g. for a given topic or for a network fragment.

Our research applies information systems paradigms to the study of social networks.

In online social networks, connected users are exposed to the content of their connections (friends in Facebook, friends and followers in Twitter, etc.) during the diffusion of information. Opinions constitute a portion of the information exchanged between users in social networks. The opinions of a friend may influence a particular user’s opinion, leading it to change over time. In addition, users generally express themselves on multiple topics. It is therefore important to discriminate to which topic the post belongs.

The focus of this thesis is the analysis of opinion change inside social networks. We propose a framework of opinion change based on design science research in information systems. The proposed framework highlights the main concepts, the model and the process of opinion change. In addition, we use topic modelling to classify content by topic or theme, sentiment analysis to extract opinions from the content. Based on the topic classification, we isolate network fragments of a user and his followees to track the opinion changes of this particular group. We also provide a generic graph database model that stores the follower graph, the metadata and content. We set criteria to define the notion of a discussion and to identify discussion patterns. The proposed definition of a discussion is composed of a series of statistical values and charts applied to each network fragment. We propose a method to extract discussion patterns and a method to compare and extract patterns from within social network datasets.

We present an instantiation of the framework of opinion change for Twitter. We show the key elements of the implementation (language and procedure) adapted for tweets analysis. We evaluated our framework using two datasets crawled on Twitter, a mixed dataset (crawled from April 16, 2016 to April 21, 2016) and a political dataset (crawled from November 28, 2016 to December 08, 2016). We have successfully applied the framework to these datasets. Regarding the topic modelling, we identify hashtag-based pooling as the best pooling method for our purpose. During the course of our research, we have identified many discussion patterns from our two datasets.

Preliminary results show that the political dataset presents more opinionated tweets than our mixed dataset. Such opinion change analysis can be useful for studying customer behavior for marketing purposes, for studying voter behavior for electoral purposes, for enriching recommendation tools and for studying information diffusion inside social networks.

Proposals for future research include: instantiations to other social networks (e.g. Instagram), real-time analysis instead of batch analysis, and deep analysis of discussions to extract recurrent patterns.

Date: Vendredi 28 juillet 2017 à 10h00

Lieu: Battelle bâtiment A - Auditoire rez-de-chaussée