MMQ-v2: Align, Denoise, and Amplify: Adaptive Behavior Mining for Semantic IDs Learning in Recommendation

TL;DR

The paper tackles the limitations of traditional ItemIDs by learning expressive semantic IDs (SIDs) that fuse multimodal item content with dynamic behavioral signals. It introduces ADA-SID, a two-fold mechanism comprising adaptive behavior-content alignment and a dynamic behavioral router within a mixture-of-quantization network to selectively amplify informative signals while suppressing noise. Empirical results on industrial and public datasets show substantial gains in both generative retrieval and discriminative ranking, with strong ablations confirming the value of each component and an online A/B showing production-level impact. The work advances robust SID learning for scalable, long-tail–aware recommender systems and highlights practical deployment benefits in real-world platforms. $R$eg$.$ $L$oss and related metrics are evaluated with established benchmarks to demonstrate effectiveness across downstream tasks and item popularity strata.

Abstract

Industrial recommender systems rely on unique Item Identifiers (ItemIDs). However, this method struggles with scalability and generalization in large, dynamic datasets that have sparse long-tail data. Content-based Semantic IDs (SIDs) address this by sharing knowledge through content quantization. However, by ignoring dynamic behavioral properties, purely content-based SIDs have limited expressive power. Existing methods attempt to incorporate behavioral information but overlook a critical distinction: unlike relatively uniform content features, user-item interactions are highly skewed and diverse, creating a vast information gap in quality and quantity between popular and long-tail items. This oversight leads to two critical limitations: (1) Noise Corruption: Indiscriminate behavior-content alignment allows collaborative noise from long-tail items to corrupt their content representations, leading to the loss of critical multimodal information. (2)Signal Obscurity: The equal-weighting scheme for SIDs fails to reflect the varying importance of different behavioral signals, making it difficult for downstream tasks to distinguish important SIDs from uninformative ones. To tackle these issues, we propose a mixture-of-quantization framework, MMQ-v2, to adaptively Align, Denoise, and Amplify multimodal information from content and behavior modalities for semantic IDs learning. The semantic IDs generated by this framework named ADA-SID. It introduces two innovations: an adaptive behavior-content alignment that is aware of information richness to shield representations from noise, and a dynamic behavioral router to amplify critical signals by applying different weights to SIDs. Extensive experiments on public and large-scale industrial datasets demonstrate ADA-SID's significant superiority in both generative and discriminative recommendation tasks.

Figures (5)

• Figure 1: Illustration of SIDs Generation Paradigm.(a) Content-based SIDs: Quantize multimodal item content into SIDs. (b) Behavior-content aligned SIDs: Incorporate collaborative signals by injecting explicit signals or aligning pre-trained representations. (c) Ours: Behavior-content shared and specific SIDs: Learn both shared and modality-specific behavior–content representations for SIDs.
• Figure 2: ADA-SID framework: We (i) use a sparse MoE-based quantization network to learn shared and modality-specific behavior–content representations, (ii) apply an adaptive behavior-content alignment mechanism that dynamically calibrates behavior–content alignment strength, and (iii) design a dynamic behavioral router that learns to assign adaptive weights to SIDs.
• Figure 3: Hyper-Parameter Analysis on Sparsity Regularization.
• Figure 4: Illustration of alignment strength controller with different hyperparameters ($\alpha$, $\beta$).
• Figure 5: Item Popularity Stratified Performance Comparison.