Source-linked AI summary
Disentangled Graph Collaborative Filtering
Xiang Wang, Hongye Jin, An Zhang, Xiangnan He, Tong Xu, Tat-Seng Chua
TL;DR
Existing collaborative-filtering representations model user-item relationships uniformly, overlooking the diverse intents behind interactions and limiting disentanglement. DGCF learns intent-aware graphs and independent intent-specific representations, with experiments demonstrating improved recommendation effectiveness, disentanglement, and interpretability.
Problem
Existing embedding functions model user-item relationships uniformly, overlooking diverse user intents and failing to disentangle intent-specific information in representations.
Method
DGCF models an intent distribution for each interaction, iteratively refines intent-aware graphs and representations, and encourages independence among intent factors.
Results
DGCF demonstrates advantages in recommendation effectiveness, disentanglement of latent user intents, and interpretability of representations across three benchmark datasets.
Takeaways & Limitations
Disentangled user intents can exploit diverse user-item relationships while supporting more interpretable representations.
Takeaways & Limitations
The authors identify future needs for side information or psychology experiments to establish ground truth on user intents and improve interpretation, alongside privacy and robustness studies.
Abstract
from arXiv · showhide
Learning informative representations of users and items from the interaction data is of crucial importance to collaborative filtering (CF). Present embedding functions exploit user-item relationships to enrich the representations, evolving from a single user-item instance to the holistic interaction graph. Nevertheless, they largely model the relationships in a uniform manner, while neglecting the diversity of user intents on adopting the items, which could be to pass time, for interest, or shopping for others like families. Such uniform approach to model user interests easily results in suboptimal representations, failing to model diverse relationships and disentangle user intents in representations. In this work, we pay special attention to user-item relationships at the finer granularity of user intents. We hence devise a new model, Disentangled Graph Collaborative Filtering (DGCF), to disentangle these factors and yield disentangled representations. Specifically, by modeling a distribution over intents for each user-item interaction, we iteratively refine the intent-aware interaction graphs and representations. Meanwhile, we encourage independence of different intents. This leads to disentangled representations, effectively distilling information pertinent to each intent. We conduct extensive experiments on three benchmark datasets, and DGCF achieves significant improvements over several state-of-the-art models like NGCF, DisenGCN, and MacridVAE. Further analyses offer insights into the advantages of DGCF on the disentanglement of user intents and interpretability of representations. Our codes are available in https://github.com/xiangwang1223/disentangled_graph_collaborative_filtering.
1 INTRODUCTION
Collaborative filtering depends on informative user and item representations, but existing methods model user-item relationships uniformly and overlook diverse latent user intents. DGCF addresses this by learning intent-aware graphs, disentangled representations, and independent intent factors.
- Collaborative filtering relies on user-item interactions to learn representations that capture user preferences.
- Prior embedding functions treat interactions as isolated instances or uniform graph edges, failing to differentiate the intents motivating different behaviors.
- Users may adopt items for distinct intents such as passing time, matching interests, or shopping for family members.
- DGCF models an intent distribution for each interaction, iteratively refining intent-aware graphs and representations while encouraging independence among intents.
- Experiments on three benchmark datasets evaluate DGCF’s recommendation effectiveness, intent disentanglement, and representation interpretability.
2 PRELIMINARY AND RELATED WORK
Collaborative filtering representation learning has progressed from isolated IDs and histories to higher-order interaction graphs, but uniform relationship modeling obscures latent intents. DGCF formulates intent-aware graphs and disentangled representations to expose finer-grained user-item relationships.
- Learning Paradigm of CF: CF predicts the likelihood that a user will adopt an item from historical user-item behavior.
- Representation Learning of CF: Interaction graphs encode collaborative signals through paths connecting users and items across multiple orders of connectivity.
- Representation Learning of CF: Representation learning evolved from independent user and item embeddings to personal histories and holistic interaction graphs.
- Limitations: Uniform relationship modeling leaves latent intents entangled, limiting interpretability and potentially producing suboptimal representations.
- Exploring User-Item Relationships: DGCF learns an intent-confidence distribution for each behavior and constructs one weighted intent-aware graph per latent intent.
- Generating Disentangled Representations: The model seeks representations composed of independent intent-specific components built from corresponding intent-aware graphs.
3 METHODOLOGY
DGCF disentangles user and item representations by coupling embedding chunks with latent intents, refining intent-aware graphs through neighbor routing and embedding propagation, and encouraging intent independence. Iterative updates produce intent-specific representations and explanatory graphs without increasing total embedding size.
- Intent-Aware Initialization: DGCF slices each user and item embedding into K intent-specific chunks and maintains a corresponding interaction graph for each latent intent.The total embedding size remains fixed rather than being doubled.
- Independence and Model Size: DGCF uses distance correlation as a regularizer to encourage independence among intent-aware representations, while retaining the same total embedding size as MF and a smaller model than NGCF.The trainable parameters are the initial user and item embeddings.
- Graph Disentangling Layer: Neighbor routing uses node-neighbor affinities to adaptively weight user-item connections before propagating information through intent-aware graphs.This differentiates the importance of relationships and refines the signals assigned to each intent.
- Iterative Update Rule: At each iteration, softmax-normalized intent scores form distributions over which intents explain a user-item interaction, after which weighted embedding propagation updates intent-specific representations.The model alternates graph refinement and representation updates for T iterations.
- Iterative Update Rule: The iterative graph update strengthens relationships between a user and historical items whose intent-specific representations are more compatible.The user’s intent-specific representation acts as a local centroid for updating edge strengths.
- Output and Interpretability: DGCF outputs disentangled representations together with intent-aware explanatory graphs that expose the information flow associated with each intent.These graphs provide evidence for what information constructs each part of the representations.
4 EXPERIMENTS
Experiments on three public datasets evaluate DGCF against multiple collaborative-filtering baselines, probe its design choices, and analyze disentanglement and interpretability. DGCF consistently improves recommendation performance, while deeper analysis links its effectiveness to intent disentanglement and interpretable intent semantics.
- Experimental Settings: Experiments use Gowalla, revised Yelp2018∗, and Amazon-Book with NGCF’s data split and recall@20 and ndcg@20 evaluation protocols.Baselines include MF, GC-MC, NGCF, DisenGCN, and MacridVAE.
- Performance Comparison: 10.88%, 4.58%, and 4.17% are DGCF’s relative recall@20 improvements over the strongest baselines on Gowalla, Yelp2018∗, and Amazon-Book, respectively.The paper reports significant improvements over all baselines across all three datasets.
- Performance Comparison: DGCF outperforms DisenGCN, while MacridVAE is the strongest baseline in most cases.The paper attributes DGCF’s advantage to diverse relationship modeling, intent-level disentanglement, and embedding propagation with neighbor routing.
- Ablation Studies: DGCF-2 substantially outperforms DGCF-1, whereas adding a third layer yields only modest further improvement.The results suggest second-order connectivity may be sufficient to distill intent-relevant information, and all DGCF depths remain superior to other baselines.
- Ablation Studies: Higher independence across intents is associated with better performance, and removing independence modeling causes larger drops in deeper DGCF models.The study measures disentanglement using distance correlation and compares DGCF with variants lacking the independence module.
- In-depth Analysis: Small contributions from individual intents can preserve near-optimal performance, but severe deficiency in any intent sharply harms results.The analysis interprets this pattern as evidence that intents are independent and cannot simply be inferred from one another.
- In-depth Analysis: Review-based analysis associates four discovered intents with interesting matching, service, price&promotion, and passing the time.Reviews for interactions sharing an intent exhibit fine-grained preferences that align with broader semantic concepts.
5 CONCLUSION AND FUTURE WORK
DGCF disentangles user intents in user and item representations by iteratively refining intent-aware interaction graphs and factorial representations. The paper concludes that disentangled intents help exploit diverse user-item relationships and interpret representations, while identifying privacy and robustness as future concerns.
- DGCF represents user-item relationships at the granularity of user intents and disentangles those intents in user and item representations.
- Its graph disentangling module iteratively refines intent-aware interaction graphs and factorial representations, while independence modeling encourages disentanglement.
- Future work includes adding side information or psychology experiments to establish ground truth on user intents, and studying privacy and robustness of factorial representations.