DisCo: Towards Harmonious Disentanglement and Collaboration between Tabular and Semantic Space for Recommendation

TL;DR

DisCo addresses the limitation of conventional recommender systems that model only the tabular space by incorporating semantic information from large language models while preserving space-specific patterns. It introduces a Dual-Side Attentive Network to capture intra- and inter-domain patterns, and regularizers—Sufficiency and Disentanglement—to maintain task-relevant information and avoid losing unique cross-space knowledge. The framework is model-agnostic and can append the generated pattern vectors to arbitrary CTR backbones, with an indexed knowledge base enabling efficient semantic retrieval. Experiments on three diverse datasets show consistent gains over both tabular-only and semantic-enhanced baselines, and ablations confirm the necessity of each component for balancing collaboration and disentanglement.

Abstract

Recommender systems play important roles in various applications such as e-commerce, social media, etc. Conventional recommendation methods usually model the collaborative signals within the tabular representation space. Despite the personalization modeling and the efficiency, the latent semantic dependencies are omitted. Methods that introduce semantics into recommendation then emerge, injecting knowledge from the semantic representation space where the general language understanding are compressed. However, existing semantic-enhanced recommendation methods focus on aligning the two spaces, during which the representations of the two spaces tend to get close while the unique patterns are discarded and not well explored. In this paper, we propose DisCo to Disentangle the unique patterns from the two representation spaces and Collaborate the two spaces for recommendation enhancement, where both the specificity and the consistency of the two spaces are captured. Concretely, we propose 1) a dual-side attentive network to capture the intra-domain patterns and the inter-domain patterns, 2) a sufficiency constraint to preserve the task-relevant information of each representation space and filter out the noise, and 3) a disentanglement constraint to avoid the model from discarding the unique information. These modules strike a balance between disentanglement and collaboration of the two representation spaces to produce informative pattern vectors, which could serve as extra features and be appended to arbitrary recommendation backbones for enhancement. Experiment results validate the superiority of our method against different models and the compatibility of DisCo over different backbones. Various ablation studies and efficiency analysis are also conducted to justify each model component.

Figures (6)

• Figure 1: Illustration of the motivation.
• Figure 2: Overview. (a) To extract the semantic knowledge, a textual description is obtained for each item using a field-value prompt template, which is then fed to a LLM for semantic embedding and stored in an indexed knowledge base. (b) The candidate item and the historical behaviors are encoded in tabular and semantic representation spaces, which are then sent to Dual-Side Attentive Network for intra-domain and inter-domain pattern vectors. The resulting pattern vectors serve as extra features and can be appended to arbitrary recommendation model. (c) Two constraints are devised to regularize the model and preserve both the aligning part and the disentangling part of useful information from the two representation spaces. The sufficiency constraint is applied on the behavior vectors and the summarized pattern vectors to preserve the useful information. The disentanglement constraint is applied on the pattern vectors from the two different domains to force the model to capture unique information from both domains.
• Figure 3: The illustration of the field-value prompt template.
• Figure 4: T-SNE visualization of the representations for the dual-side attentive network output (ML-1M).
• Figure 5: T-SNE visualization of the representations for the dual-side attentive network output (AZ-Toys).