A Machine Learning based Approach to Identify User Interests from Social Data

Abstract

Social media platforms like Twitter, Facebook, Instagram, etc., are considered a common source of extracting information about individuals, such as their needs, interests, and opinions. Our major contribution in this paper is to identify user interests and desires related to the fashion industry in Pakistan. Since people in Pakistan mostly write tweets and reviews in Roman Urdu, the dataset we focused on in this research was comprised of Roman Urdu Tweets and Google Map reviews. From the literature, we observed that not much effort has been done on Roman Urdu tweets and reviews because of its being a low resource language. In terms of methodology, we applied LDA, LSA, and BERT for topic modeling; Vadar combined with TextBlob and DistilBert for sentiment analysis; and K-Means for identifying user clusters with similar interests. In our experiments, we used 15000 tweets and 6000 Google reviews. We were able to create five distinct clusters for each brand. These clusters were further used to track the users based on their interests. We evaluated the performance of our approach and validated it empirically based on Cohen's Kappa score, and achieved a score of 0.45 that shows moderate agreement between human and machine.