Xinyi Zeng, Wen Shi, Yang Bai, Li Jin, Xiaoyu Li, Zhi Guo
Session-based recommendation, multi-modal, PID, cold start
Session-based recommendation (SBR) predicts the next item by analysing user’s short interaction sequences. The increasing popularity of Internet and mobile applications enables the deployment of session-based recommendation algorithms in terminal systems to fulfill user’s personalised and real-time requirements. In practical scenarios, data from terminal systems is frequently multimodal. We delve into session-based recommendation within multimodality and propose a graph collaborative guided session- based recommendation method, namely GCG. GCG aims to enhance the alignment and fusion of multimodal data and constructing more robust session features. First, we include modality alignment guided by collaborative information from graph neural networks. After modality alignment, we propose multiple stars representations in transformers to fuse collaborative, textual, and visual information from diverse data sources, overcoming the limitations of modal monopoly. Subsequently, we demonstrate the applicability of proportion–integral–differential (PID) in session- based recommendation and introduce session embeddings derived from PID, which model a more robust session representation incorporating constant, integral, and differential terms. Extensive experiments on three real datasets demonstrate the effectiveness and advancement of the proposed GCG. Further analysis also proves that GCG can effectively alleviate the cold start problem in SBR.
Important Links:
Go Back
