Source-linked AI summary
DisenHAN: Disentangled Heterogeneous Graph Attention Network for Recommendation
Yifan Wang, Suyao Tang, Yuntong Lei, Weiping Song, Sheng Wang, Ming Zhang
TL;DR
Existing heterogeneous graph recommenders overlook disentangled latent factors and the semantic influence of meta paths. DisenHAN decomposes high-order connectivity into meta relations and propagates aspect-specific information, consistently outperforming state-of-the-art models while supporting interpretable representations.
Problem
Existing heterogeneous graph neural networks entangle latent factors from different aspects and simplify meta paths, overlooking their richer semantic information.
Method
DisenHAN learns disentangled user and item representations by decomposing high-order connectivity into meta relations and iteratively identifying their major aspects during embedding propagation.
Results
DisenHAN consistently outperforms existing state-of-the-art models on three real-world datasets and shows effectiveness and interpretability in recommendation.
Takeaways & Limitations
Aspect-specific propagation enables DisenHAN to aggregate semantic information from automatically generated meta paths for top-N recommendation.
Takeaways & Limitations
Future work will examine dynamic user-item interactions in temporal graphs and extend DisenHAN to node classification and clustering in HINs.
Abstract
from arXiv · showhide
Heterogeneous information network has been widely used to alleviate sparsity and cold start problems in recommender systems since it can model rich context information in user-item interactions. Graph neural network is able to encode this rich context information through propagation on the graph. However, existing heterogeneous graph neural networks neglect entanglement of the latent factors stemming from different aspects. Moreover, meta paths in existing approaches are simplified as connecting paths or side information between node pairs, overlooking the rich semantic information in the paths. In this paper, we propose a novel disentangled heterogeneous graph attention network DisenHAN for top-$N$ recommendation, which learns disentangled user/item representations from different aspects in a heterogeneous information network. In particular, we use meta relations to decompose high-order connectivity between node pairs and propose a disentangled embedding propagation layer which can iteratively identify the major aspect of meta relations. Our model aggregates corresponding aspect features from each meta relation for the target user/item. With different layers of embedding propagation, DisenHAN is able to explicitly capture the collaborative filtering effect semantically. Extensive experiments on three real-world datasets show that DisenHAN consistently outperforms state-of-the-art approaches. We further demonstrate the effectiveness and interpretability of the learned disentangled representations via insightful case studies and visualization.
1 INTRODUCTION
Existing recommendation methods struggle with sparse interactions and cold start, while heterogeneous graph approaches often entangle factors and simplify meta-path semantics. DisenHAN addresses these issues by learning disentangled aspect representations and semantically propagating information through automatically generated meta paths.
- Collaborative filtering methods suffer from sparse user-item interactions and cold-start problems, motivating the use of contextual information.
- Heterogeneous information networks model users, items, and contextual entities with multiple node and relation types, using meta paths to capture semantic relatedness.
- Existing approaches often entangle latent factors and treat meta paths as generic side information or connecting paths, overlooking aspect-specific semantic influence.
- DisenHAN learns disentangled user and item representations by identifying the major aspect of relation information flows through iterative semantic embedding propagation.
- Stacked propagation layers aggregate information from multiple meta relations, while experiments report improved performance and interpretability over existing state-of-the-art models.
2 RELATED WORK
Prior work applies heterogeneous networks, graph neural networks, and disentangled representation learning to recommendation, but important semantic distinctions remain incompletely modeled. HIN-based GNNs improve high-order collaborative modeling while still overlooking different semantic aspects between node pairs.
- HIN-based collaborative filtering methods use context-rich meta paths or knowledge-aware embeddings, but some rely heavily on domain-expert prior knowledge.
- Graph neural networks aggregate neighbor features end to end and can model high-order connectivity more explicitly than traditional graph embeddings.
- Recent HIN-based GNNs incorporate heterogeneous information but still directly aggregate different node types or overlook distinct semantic aspects between node pairs.
- Disentangled representation learning separates explanatory factors and has been associated with robustness, generalization, and interpretability in downstream processes.
3 PROBLEM DEFINITION
The paper formalizes heterogeneous information networks, meta paths, and meta relations for top-N recommendation with contextual user-item interactions. It replaces manually selected meta paths with combinations of individual semantic relations and predicts engagement from aspect-specific representations.
- A heterogeneous information network maps nodes and edges to multiple types, requiring |A| + |R| > 2.
- A meta path is a composite relation between node types, but selecting dataset-specific meta paths manually becomes intractable as their number grows.
- A meta relation is the typed semantic relation ⟨ϕ(s), ψ(e), ϕ(t)⟩ for an edge from source node s to target node t.
- The recommendation problem uses binary user-item interactions plus contextual information such as social relationships, brands, and categories.
- The model predicts the probability that user u will engage with item v from their aspect-specific representations, graph G, interactions Y, and parameters Θ.
4 THE PROPOSED MODEL
DisenHAN enriches heterogeneous user and item representations by grouping neighbors by meta relation, separating features into aspect subspaces, and iteratively selecting dominant semantic aspects. Stacked layers propagate multi-hop information for matching-score prediction.
- DisenHAN is designed for HIN recommendation because node representations arise from different aspects determined by corresponding meta relations.
- For each target node, neighbors are grouped by meta relation and their attributes are projected into different subspaces before aspect-specific aggregation.
- The propagation mechanism iteratively identifies the major aspect of each meta relation and produces combined semantic information for the target node.
- Stacking L DisenHAN layers aggregates multiple hops of neighbor information to enrich user and item representations used for matching-score prediction.
4.2 Disentangled Content Transformation
DisenHAN transforms heterogeneous node features into aspect-specific subspaces before propagation. It uses multiple channels for disentangled representations and reduces the number of aspects across increasingly distant hops.
- Type-specific transformations project target and source features before they enter the DisenHAN layer.
- Target-node features are projected into K different subspaces to represent distinct aspects.
- The model outputs three disentangled representation channels for a target node and feeds them back as features for the next iteration.
- The number of aspects is constrained to decrease as propagation reaches more distant neighbor hops.
4.3 Disentangled Propagation Layer
The disentangled propagation layer first aggregates neighbors within each meta relation and aspect, then combines meta relations by iteratively identifying their major aspects.
- Intra-relation Aggregation: The intra-relation layer aggregates grouped source-node features to capture semantic information for each meta relation.
- Intra-relation Aggregation: Attention weights learn unequal source-node contributions under different aspects.
- Intra-relation Aggregation: The attention weight for a meta relation is semantic-specific and captures its corresponding semantic information.
- Inter-relation Aggregation: Inter-relation aggregation assigns each meta relation a nonnegative probability of belonging to each target-node factor.
- Inter-relation Aggregation: The model iteratively searches meta relations for their major aspects and feeds the resulting aspect representations back into neighbor attention.
- Iterative Forward Propagation: Algorithm 1 initializes K channels from target-node features and propagates grouped-neighbor information across iterations.
4.4 Recommendation
After L propagation layers, DisenHAN represents users and items with aspect-specific embedding vectors and scores recommendations by summing aspect-wise inner products.
- L propagation layers produce K embedding vectors for each user and item, with each vector representing a different aspect.
- The recommendation score sums the inner products between corresponding user and item aspect embeddings.
4.5 Training
DisenHAN trains on implicit-feedback recommendation as binary classification with negative sampling and mini-batch optimization.
- The model treats implicit-feedback top-N recommendation as a binary classification problem whose score estimates future item selection likelihood.
- Training uses positive instances and uniformly sampled negative instances from unobserved interactions.
- Mini-batch training pads source-node sets to a common maximum, while fixed-number neighbor sampling addresses long-tail relation sizes.
4.6 Analysis of DisenHAN
DisenHAN’s iterative propagation identifies which meta relations are relevant to each latent aspect by clustering grouped source neighbors, then aggregates aspect-specific information across layers. The process has an EM interpretation with monotonic log-likelihood improvement and supports soft selection of multi-hop meta paths.
- Iterative Forward Propagation Process: The propagation process combines features from related meta relations to enrich a target node’s aspect k.It infers related source neighbors by identifying those that form a cluster in aspect k’s subspace.
- Convergence Analysis: DisenHAN models iterative forward propagation as an expectation maximization algorithm for a mixture model.The target representation is treated as the latent variable, while neighboring representations and meta-relation weights provide observations and relation-specific quantities.
- Convergence Analysis: The EM updates monotonically improve the log-likelihood, so the iterative propagation algorithm converges.The E step sets q(R) to p(R|C;Z), and the M step maximizes the ELBO with respect to Z.
- Automatic Meta Paths Selection: Stacking propagation layers lets DisenHAN aggregate corresponding aspect information from multiple hops, providing a soft way to select meta paths.Each layer incorporates one-hop neighbor information while iteratively identifying aspect weights for meta relations.
5 EXPERIMENTS
Experiments evaluate DisenHAN on three public recommendation datasets against classical CF, HIN-based, and GNN-based baselines, including sparsity, component, and interpretability analyses. DisenHAN consistently performs best, with gains especially evident under sparse interactions and when modeling multiple semantic aspects.
- Experimental setup: Experiments use Yelp, Amazon-electricity, and MovieLens for top-N recommendation, comparing classical CF, HIN-based, and GNN-based methods.The evaluation uses chronological train/validation/test splits and Precision, Recall, and NDCG metrics.
- Overall comparison: DisenHAN consistently outperforms other baselines across all datasets, improving Prec@10, Recall@10, and NDCG@10 over the strongest baselines.The reported improvements are 2.3%, 3.5%, and 2.1% for Prec@10; 3.6%, 10.8%, and 3.0% for Recall@10; and 4.6%, 9.0%, and 2.1% for NDCG@10 on Yelp, Amazon-electricity, and MovieLens, respectively.
- Disentangled representation: Increasing the number of disentangled aspects improves recommendation performance, with the best performance around 5 aspects and lower gains when K > 10.The authors associate very large aspect counts with an overly complex semantic structure.
- Propagation design: More iterations generally improve performance before saturation, while stacking two propagation layers improves performance over one and deeper stacking can introduce overfitting.Performance becomes stable at about five iterations; more than two propagation layers provide small neighbor influence and may introduce noise.
- Interpretability: Visualization indicates that DisenHAN separates users and items into distinct groups and produces discernible clustering associated with different users.The case study also reports different aspects being associated with distinct meta relations across propagation layers.
6 CONCLUSION AND FUTURE WORK
DisenHAN uses disentangled semantic propagation over meta relations for top-N recommendation in heterogeneous information networks. Experiments on three real-world datasets support its effectiveness and interpretability, while dynamic temporal interactions remain future work.
- DisenHAN uses meta relations to decompose high-order connectivity and aggregate aspect-specific semantic information for users and items.Its disentangled embedding propagation layer identifies major aspects of information flows and supports automatic meta-path generation.
- Extensive experiments on three real-world datasets demonstrate DisenHAN’s effectiveness and interpretability.
- Future work will examine whether DisenHAN can capture dynamic user-item interactions in temporal graphs.