Semantic Codebook Learning for Dynamic Recommendation Models

TL;DR

This work tackles dynamic sequential recommendation under sparse, noisy user-item interactions by reducing the large parameter search space. It introduces SOLID, a framework that disentangles parameter generation through semantic sequences, a semantic metacode, and a semantic codebook, along with a dual trunk that fuses homogeneous (semantic) and personalized (item) signals. SOLID comprises Semantic Parameter Generation, Semantic Metacode Learning, and Semantic Codebook Learning to enable robust, scalable adaptive parameters. Empirical results across eight datasets show SOLID consistently outperforms existing DSR methods with improved stability and robustness, highlighting the value of multimodal semantic guidance for personalized, real-time recommendations.

Abstract

Dynamic sequential recommendation (DSR) can generate model parameters based on user behavior to improve the personalization of sequential recommendation under various user preferences. However, it faces the challenges of large parameter search space and sparse and noisy user-item interactions, which reduces the applicability of the generated model parameters. The Semantic Codebook Learning for Dynamic Recommendation Models (SOLID) framework presents a significant advancement in DSR by effectively tackling these challenges. By transforming item sequences into semantic sequences and employing a dual parameter model, SOLID compresses the parameter generation search space and leverages homogeneity within the recommendation system. The introduction of the semantic metacode and semantic codebook, which stores disentangled item representations, ensures robust and accurate parameter generation. Extensive experiments demonstrates that SOLID consistently outperforms existing DSR, delivering more accurate, stable, and robust recommendations.

Figures (5)

• Figure 1: (a) describes multimodal user behavior data that includes images, text, and IDs. (b) describes the forward propagation of DSR, which is divided into two pathways: the first pathway processes user behavior data composed of IDs through a parameter generator to produce the parameters for the dynamic layers of the primary model. The second pathway processes the same ID-based user behavior data through the primary model's static layer, then through the dynamic layer, resulting in the prediction output. (c) and (d) compare the parameter generation patterns of existing DSR and SOLID. (e) compares the performance of our method and SR models and DSR Models on four multi modal recommendation datasets and four single modal recommendation datasets. The results show that our method significantly enhances performance on extensive datasets.
• Figure 2: The framework of the SOLID, which consists of three main modules: Semantic Parameter Generation (SPG), Semantic Metacode Learning (SML), and Semantic Codebook Learning (SCL). SPG first converts item representations into semantics and constructs a semantic sequence to generate parameters in a structured manner. Subsequently, SML generates model parameters based on both the item sequence and the semantic sequence, and it jointly trains the model, accommodating both homogeneous and heterogeneous information. More importantly, the semantic encoder it learns can be transformed into metacode, which then provides a good initial value for the codebook. Finally, SCL learns a semantic codebook to improve the process of the parameter generation. Among them, $\mathcal{L}_{\text{Rec}}=l_{\text{CE}}(y, \hat{y}), \mathcal{L}_{\text{Con}} = l_{\text{MSE}}(\mathbf{E}_v, \mathbf{E}_v')$.
• Figure 3: UAUC comparison of the proposed method and baseline on the CDs and Electronic datasets.
• Figure 4: UAUC comparison of the proposed method and baseline on the Book and Music datasets.
• Figure 5: Hyperparameter Grid Search.