Etudiant : OULAJA SAFAE

Filière : Master Big Data Analytics & Smart Systems (BDSaS)

Encadrant : Pr. NFAOUI EL HABIB

Annèe : 2025

Résumé : In e-commerce, enhancing user experience relies heavily on effective recommender systems. Session based recommendation, which focuses on predicting a user’s next action based solely on their short-term browsing behavior, presents significant challenges due to the lack of long-term user profiles and the dynamic nature of session interactions. This thesis introduces DualEmbedRecommender, a novel recommendation architecture built upon a Transformer-based backbone, enhanced with a dual embedding strategy. The model leverages self-attention mechanisms to capture intra-session dependencies and employs a sliding-window approach for data augmentation, alongside regularization techniques such as label smoothing to improve generalization. To further extend its capabilities, the model is evaluated not only on the anonymized Diginetica dataset but also on the Amazon All Beauty dataset, which provides rich semantic metadata. In this extended setting, we integrate LLM-based embeddings, generated using pretrained language models, to enrich item representations. Experiments demonstrate that these embeddings significantly boost performance, validating the synergy between deep sequential modeling and semantic item understanding. The proposed approach achieves state-of-the-art results on standard metrics such as MRR@20 and HR@20, showing its effectiveness across both sparse and semantically rich environments.