Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Dual-Perspective Disentangled Multi-Intent Alignment for Enhanced Collaborative Filtering

Shanfan Zhang, Yongyi Lin, Yuan Rao, Bingcan Xia, Tingting Xin, Chenlong Zhang

Abstract

Personalized recommendation requires capturing the complex latent intents underlying user-item interactions. Existing structural models, however, often fail to preserve perspective-dependent interaction semantics and provide only indirect supervision for aligning user and item intents, lacking explicit interaction-level constraints. This entangles heterogeneous interaction signals, leading to semantic ambiguity, reduced robustness under sparse interactions, and limited interpretability. To address these issues, we propose DMICF, a Dual-Perspective Disentangled Multi-Intent framework for collaborative filtering. DMICF models interactions from complementary user- and item-centric perspectives and employs a macro-micro prototype-aware variational encoder to disentangle fine-grained latent intents. Interaction-level supervision enforces dimension-wise alignment between user and item intents, grounding latent factors and enabling their collaborative emergence. Importantly, each component is architecturally flexible, and performance is robust to specific module instantiations. We offer a theoretical analysis to help explain how prototype-aware conditioning may alleviate posterior collapse, while the reconstruction objective promotes intent-wise contrastive alignment between positive and negative interactions. Extensive experiments on multiple benchmarks demonstrate consistent improvements over strong baselines, with ablations validating each core component.

Dual-Perspective Disentangled Multi-Intent Alignment for Enhanced Collaborative Filtering

Abstract

Personalized recommendation requires capturing the complex latent intents underlying user-item interactions. Existing structural models, however, often fail to preserve perspective-dependent interaction semantics and provide only indirect supervision for aligning user and item intents, lacking explicit interaction-level constraints. This entangles heterogeneous interaction signals, leading to semantic ambiguity, reduced robustness under sparse interactions, and limited interpretability. To address these issues, we propose DMICF, a Dual-Perspective Disentangled Multi-Intent framework for collaborative filtering. DMICF models interactions from complementary user- and item-centric perspectives and employs a macro-micro prototype-aware variational encoder to disentangle fine-grained latent intents. Interaction-level supervision enforces dimension-wise alignment between user and item intents, grounding latent factors and enabling their collaborative emergence. Importantly, each component is architecturally flexible, and performance is robust to specific module instantiations. We offer a theoretical analysis to help explain how prototype-aware conditioning may alleviate posterior collapse, while the reconstruction objective promotes intent-wise contrastive alignment between positive and negative interactions. Extensive experiments on multiple benchmarks demonstrate consistent improvements over strong baselines, with ablations validating each core component.

Paper Structure

This paper contains 29 sections, 1 theorem, 22 equations, 8 figures, 4 tables.

Table of Contents

  1. INTRODUCTION
  2. METHODOLOGY
  3. Model
  4. Dual-perspective structural encoding
  5. Prototype-aware variational intent generation
  6. Interaction-driven semantic alignment
  7. Model Optimization
  8. Model Analysis
  9. Time Complexity Analysis
  10. Mitigating Intent Collapse.
  11. Optimization Analysis of Intent Alignment
  12. Comparative Implications.
  13. EXPERIMENTS
  14. Experimental Setting
  15. Datasets
  16. ...and 14 more sections

Key Result

Proposition 1

[Aggregate Posterior as Implicit Mixture] Let $\mathbf{o}^{\left ( u \right ) }_{i}$ denote the latent intent variable of user $u_{i}$. The variational posterior $q_{\phi}(\mathbf{o}^{\left ( u \right ) }_{i} | \textbf{T}^{\left ( u \right ) }_{i})$ captures individual-level intent uncertainty, whil where $\mathbb{P}(\mathcal{C}^{(u)}_k \mid u_{i})$ denotes the soft assignment of user $u_{i}$ to p

Figures (8)

  • Figure 1: Overview of DMICF, which preserves dual-perspective semantics and enables fine-grained intent alignment.
  • Figure 2: Per-epoch training time of DMICF versus baselines.
  • Figure 3: Early-stage performance (first 30 epochs) of DMICF, IPCCF, and LightCCF. Slower models are omitted for clarity.
  • Figure 4: Performance of DMICF, IPCCF and BIGCF across user groups binned by interaction frequency. Tmall spans $\left [ 0,20 \right )$, and ML-10M has no users in $\left [ 0,10 \right )$.
  • Figure 5: Hyperparameter study of DMICF.
  • ...and 3 more figures

Theorems & Definitions (2)

  • Remark 1: Optimization Insight.
  • Proposition 1