Disentangled Contrastive Collaborative Filtering

TL;DR

This work addresses sparsity and robustness in graph-contrastive collaborative filtering by introducing DCCF, a framework that combines disentangled intent representations with adaptive, globally informed augmentation. It defines $K$ latent intents with global prototypes $(\mathbf{c}_u^k, \mathbf{c}_v^k)$ and computes intent-aware representations that are fed into a graph neural network, alongside learnable masks that adaptively perturb the interaction graph for SSL. A cross-view contrastive objective, built from three augmented views and fused with a BPR loss, enables robust, intent-aware self-supervision, reducing noise amplification and mitigating over-smoothing. Empirical results on Gowalla, Amazon-book, and Tmall demonstrate state-of-the-art performance and improved robustness to data sparsity, with ablations validating the contributions of disentangled encoding, adaptive masking, and SSL components; code is released for reproducibility.

Abstract

Recent studies show that graph neural networks (GNNs) are prevalent to model high-order relationships for collaborative filtering (CF). Towards this research line, graph contrastive learning (GCL) has exhibited powerful performance in addressing the supervision label shortage issue by learning augmented user and item representations. While many of them show their effectiveness, two key questions still remain unexplored: i) Most existing GCL-based CF models are still limited by ignoring the fact that user-item interaction behaviors are often driven by diverse latent intent factors (e.g., shopping for family party, preferred color or brand of products); ii) Their introduced non-adaptive augmentation techniques are vulnerable to noisy information, which raises concerns about the model's robustness and the risk of incorporating misleading self-supervised signals. In light of these limitations, we propose a Disentangled Contrastive Collaborative Filtering framework (DCCF) to realize intent disentanglement with self-supervised augmentation in an adaptive fashion. With the learned disentangled representations with global context, our DCCF is able to not only distill finer-grained latent factors from the entangled self-supervision signals but also alleviate the augmentation-induced noise. Finally, the cross-view contrastive learning task is introduced to enable adaptive augmentation with our parameterized interaction mask generator. Experiments on various public datasets demonstrate the superiority of our method compared to existing solutions. Our model implementation is released at the link https://github.com/HKUDS/DCCF.

Figures (6)

• Figure 1: The overall framework of our proposed DCCF model involves adaptive augmentation through the integration of global intent disentanglement and interaction pattern encoding, resulting in disentangled environment-invariant representations.
• Figure 2: Performance comparison w.r.t. data sparsity over different user/item groups on Gowalla data.
• Figure 3: Performance w.r.t the number of latent intents.
• Figure 4: Distribution of latent intent prototypes.
• Figure 5: Case study of intent-aware global user relations. Non-locally connected users ($u_{1155}$ and $u_{32856}$) can be identified with similar user preference (large item category overlap) via our learned disentangled representations.