Scalable Losses in Session-based Recommendation Systems with Deep Learning Architectures

• Ano de defesa: 2023
• Autor(a) principal: Veiga, Tobias Mesquita Silva da
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Aprendizagem profunda; Attention mechanism; Deep learning; escalabilidade; Graph leural networks; Mecanismo de atenção; Recomendações baseadas em sessões; Redes neurais de grafos; scalability; Session-based recommendation
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/55/55134/tde-03072023-135346/

In recommendation systems the objective is to improve user experience by suggesting interesting contents and quickly providing what they look for. To help with this task, Deep learning has proven to be an efficient tool, especially in recommendation systems where user navigation sessions are anonymous. The nature of Deep Learning tools such as Recurrent Networks, Graph Neural Networks and Attention Mechanism makes them capable of dealing with variable length data in large scale, which is ideal for processing user sessions. In this context, many state of the art Deep Learning models have the limitation of not being scalable to large datasets, where the number of unique items to recommend is very large. In this work we explore how scalable loss functions can modify the results of previous works and we introduce a new challenging dataset to assess whether such modifications really make the models scalable.