Exploiting Inter-session Dynamics for Long Intra-Session Sequences of Interactions with Deep Reinforcement Learning for Session-Aware Recommendation

• Autor(a) principal: Ticona, Gustavo Junior Escobedo
• Data de Publicação: 2021
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Biblioteca Digital de Teses e Dissertações da USP
• Texto Completo: https://www.teses.usp.br/teses/disponiveis/55/55134/tde-23062021-105306/
• topic: Aprendizado por reforço profundo; Deep learning; Recomendação ciente de sessão; Recommender systems; Redes neurais recorrentes hierarquicas; Reinforcement learning; Session-aware recommendation; Sistemas de recomendação

Recommender systems are tools whose objective is to filter relevant content to users according to their preferences. Recently, due to the new demands of electronic business where most of users are not authenticated, Session-based recommender systems emerged. This approach models session data (e.g. sequences of interactions, item metadata) to predict which items will be relevant for the user during the current session. Session-aware approaches include representations from users past sessions to improve performance on fresh new sessions. However, current approaches only exploit these representations at the beginning of the session which in a long sequence of interactions does not take advantage of possible changes of interest during the same session. Consequently, in this research work, we explore the possibility of exploiting inter-session representations to improve recommendation performance. We proposed an adaptation of the Deep Deterministic Policy Gradient algorithm on a session-aware recommender model to train a policy that handles the interaction between the current intra-session state and inter-session representations. We performed several experiments on two datasets from different domains finding key factors that affect session-aware models performance. However, we could not find strong evidence to claim that inter-session dynamics can improve performance during long sequences of intra-session interactions.