Semantic Codebook Learning for Dynamic Recommendation Models

Zheqi Lv, Shaoxuan He, Tianyu Zhan, Shengyu Zhang, Wenqiao Zhang, Jingyuan Chen, Zhou Zhao, Fei Wu

Abstract:

Dynamic Sequential Recommendation (DSR) systems, which adapt to users' changing preferences by adjusting their parameters in real time, have emerged as a significant advancement over traditional static models. Despite their advantages, DSR systems face challenges including large item parameter spaces and heterogeneous user-item interactions, which can destabilize the recommendation process. To address these issues, we introduce the Semantic Codebook Learning for Dynamic Recommendation Models (SOLID) framework. SOLID compresses the parameter generation model's search space and utilizes homogeneity within the recommendation system more effectively. This is achieved by transforming item sequences into semantic sequences and employing a dual parameter model, which combines semantic and item-based cues to tailor recommendation parameters. Our innovative approach also includes the creation of a semantic codebook, which stores disentangled item representations to ensure stability and accuracy in parameter generation. Through extensive testing, SOLID has shown to surpass traditional DSR systems, providing more precise, stable, and dynamically adaptable recommendations.~\footnote{Our source code can be referred to \url{https://anonymous.4open.science/r/SOLID-0324}}.

Primary Subject Area: [Engagement] Multimedia Search and Recommendation

Secondary Subject Area: [Content] Multimodal Fusion

Relevance To Conference: This study breaks through the performance bottlenecks of existing dynamic recommendation models, which are limited by reliance on single-modal input and unconstrained parameter search spaces, by leveraging multimodal data to decouple the parameter generation process of dynamic recommendation models. This research not only advances the development of multimodal recommendation systems but also provides new perspectives and methods for parameter optimization in dynamic recommendation systems, holding significant theoretical value and practical application prospects.

Submission Number: 108