MSc. Thesis Defense: Giray Havur

IDENTIFICATION OF USER INTERESTS IN SOCIAL MEDIA

Giray Havur
Computer Science & Engineering, M.Sc. Thesis, 2014

Thesis Jury
Assoc. Prof. Yücel Saygın (Thesis Supervisor), Assoc. Prof. Berrin Yanıkoğlu, Assoc. Prof. Ali Koşar

Date &Time: July 22nd 2014 – 15:30

Place: FENS 2072

Keywords: Online communities, information retrieval, semantic analysis, sentiment analysis, social networks

Abstract

Social media has taken an important part in our lives in a short amount of time. People share parts of their experiences, opinions, and interests with others in a timely-fashion on these platforms. In recent years, fast growth of user population in social media is not only driving the research towards analyzing its inhabitants for fulfilling their expectations but also making it a very crucial information source for decision making processes in societies and in businesses. In this work, we propose methods for identifying users and their interests by using the multimedia data shared in social media. We show effectiveness of these methods in three applications. Our first application considers extracting political interests of Turkish Twitter users. We collect tweets that include a set of predefined words representing two different political stances in Turkey. We extract profile images of the users who wrote those tweets and apply a computer vision technique called image context extraction on this set of images to obtain some textual explanations for each picture. The main goal of this work is inferring proportions of two different political stances to forecast results of March 2014 local elections. Our results show that the proportions obtained from our method are almost the same as the vote percentages of two parties. Our second application matches people whose identification information (i.e. Name and location) is provided, to their Facebook profiles. Our method achieve matching of 1332 profiles out of 1500 people. The third application deals with building a community about a given topic representing an interest by condensing existing communities in a social media platform. We collect members of the communities about the given topic in a set and apply our relevance scoring method on these members. Those who receive a score below a threshold value are assumed to be irrelevant to given topic and they are eliminated so that remaining users in the set are the ones relevant to given topic. We validated the results of our framework by a user-study. There is a 76\% of match between user labelled and automated results.