The School of Computer Science is pleased to present…
Date: Monday, May 12th, 2025
Time: 10:00 AM
Location: Essex Hall, Room 122
Recent advances in natural language processing and a growing number of user reviews on online platforms enable more accurate recommendations by extracting and analyzing personal preferences from review texts. Despite decades of progress, many existing review-based systems treat each review as a whole block of text, overlooking finer-grained, aspect-level sentiments that reveal precisely what users like and dislike. This approach often results in suboptimal recommendations, particularly when reviews vary widely in focus or when only certain segments carry actionable insights.
In this thesis, we enhance the performance of review-based recommender systems by integrating both explicit and implicit user–item interactions derived from profiles built on aspect-level sentiments. These profiles are constructed from sentiments expressed in reviews about domain-specific aspects. To achieve this, we leverage DeBERTa (Decoding-enhanced BERT with disentangled attention) for aspect-based sentiment analysis, capturing user preferences from past reviews and item characteristics from public opinion.
Our experiments demonstrate that our model outperforms some of the existing review-based methods by performing fine-grained analysis of reviews, focusing on the most informative segments of the reviews and their associated sentiments, to build robust user and item profiles. This profile construction reduces the system’s reliance on review text, an independence that is particularly valuable in real-world scenarios where predictions must be made for unseen user–item interactions without available reviews.
Internal Reader: Dr. Muhammad Asaduzzaman
External Reader: Dr. Bharat Maheshwari
Advisor: Dr. Pooya Moradian Zadeh
Chair: Dr. Sherif Saad