Identifying and Profiling User Interest over time using Social Data

With immense population growth in recent years, social data is growing at a rapid pace, which in turn can prove to be a rich source of hidden information. This work focuses on identifying user interest in electronic products, especially smartphones, using social data. This will help electronic businesses in the personalized marketing of their products. From the literature, most of the existing approaches attempted to identify user interest based on their ratings. In our understanding, the contents of reviews are equally important in identifying people's interests. Therefore, in this paper, we proposed a framework that identifies user interests based on their reviews and their ratings. Moreover, it performs an analysis of the aforementioned reviews, and profiles user interest. To achieve this, we used website data, written in the Roman Urdu language. To the best of our knowledge, very limited research has been carried out on the Roman Urdu dataset, as it is considered a low-resource language. Concerning our methodology, we first performed topic modeling using Latent Dirichlet Allocation (LDA), Bidirectional Encoder Representations from Transformers (BERT), and a hybrid of both. Based on the identified topics, we performed user interest profiling based on the probabilities of each model/brand using the Top2Vec model. We compared our results of topic modeling using reviews and reviews plus ratings. For topic modeling, we measure coherence score which we observe 52% for the hybrid approach while 47% and 45% for “BERT” and “LDA” respectively. Finally, For topic modeling, we perform human-based validation by comparing human-identified topics with the ones identified by our model.