Reproduce “Multi-Behavior Generative Recommendation” on OpenResearch

About this paper

Zihan Liu, Yupeng Hou, Julian McAuley

Researchers at UCSD developed MBGen, a generative framework for multi-behavior sequential recommendation that explicitly predicts both the next behavior type and the next item in a two-step process, achieving 30-70% performance improvements over prior state-of-the-art models on public datasets.

Key insights

• Reframing multi-behavior sequential recommendation as a generative task that predicts both behavior and item sequentially enables a more holistic understanding of user intent and facilitates novel platform functionalities.
• Tokenizing items into multiple balanced discrete tokens, rather than using a single item ID, is crucial for effectively managing the vast item catalog and improving learning efficiency in generative recommendation models.
• Integrating behavior type information at multiple levels within a Transformer (via behavior tokens in the sequence, position-routed experts, and behavior-injected FFNs) is vital for capturing fine-grained behavior-item dependencies and scaling model capacity effectively.

Summary by alphaXiv.