Interest representations in deep news recommender systems

Abstract

News recommender systems (NRS) aim to reduce information overload by suggesting articles tailored to user interests. However, traditional systems often rely on single-vector user representations. These may fail to capture the full diversity of user preferences. This thesis introduces a novel approach using multi-interest user representations which is combined with disentanglement objective to ensure distinct and non-overlapping interest vectors. The study includes a review of both traditional and deep learning-based news recommendation methods, followed by the development of new multi-interest model. This model is tested on the MIND dataset, a large collection of user behavior logs from Microsoft News. The evaluation focuses on enhancing click prediction accuracy, achieving clear and separate interest representations, improving recommendation fairness and diversity, and reducing bias through geometric analysis of embeddings. Results demonstrate that multi-interest user representations enhance click prediction accuracy and produce more balanced recommendations. Disentanglement techniques reduce overlap between interest vectors, creating clearer user profiles. However, the approach may reduce recommendation diversity, suggesting a need for careful tuning. This framework offers the potential for more personalized, fair, and unbiased news recommendations across various domains.