Reproduce “Image Fusion for Cross-Domain Sequential Recommendation” on OpenResearch

About this paper

Wangyu Wu, Siqi Song, Xianglin Qiu, Xiaowei Huang, Fei Ma, Jimin Xiao

A new method, Image Fusion for Cross-Domain Sequential Recommendation (IFCDSR), from Xi'an Jiaotong-Liverpool University and The University of Liverpool, enhances cross-domain sequential recommendation by integrating image features via a frozen CLIP model and employing a multi-attention mechanism. It consistently outperformed existing state-of-the-art models on Amazon datasets, showing an additional 0.6% MRR improvement from image fusion and 0.45% from its multiple attention strategy.

Key insights

• Visual features derived from pre-trained vision-language models significantly enrich item representations, leading to a more comprehensive understanding of user preferences in sequential recommendation.
• Employing a multiple attention layer that processes intra-sequence, inter-sequence, and multimodal information separately helps mitigate domain bias and captures diverse user preferences effectively.
• Fusing predictions from both ID-based and image-based features, as well as from different domain sequences, yields more accurate and robust cross-domain recommendations.

Summary by alphaXiv.