HGCL: Hierarchical Graph Contrastive Learning for User-Item Recommendation

TL;DR

HGCL addresses the gap in graph contrastive learning for recommendations by explicitly modeling hierarchical item structures. It introduces a two-stage framework: (1) cross-layer contrastive pre-training on a user-item graph with noise-augmented embeddings, and (2) item embedding compression via t-SNE followed by deterministic $\rho \times \theta$ sector clustering to build a two-hierarchy bipartite graph for fine-tuning. Empirical results on Yelp2018, Amazon-Kindle, and Alibaba-iFashion show that HGCL yields higher Recall@20 and NDCG@20 than state-of-the-art baselines, validating the value of multi-resolution item signals. The approach offers a practical, privacy-friendly path to incorporate hierarchical item relations into GCL-based recommendations, with future work extending to large-scale industrial datasets.

Abstract

Graph Contrastive Learning (GCL), which fuses graph neural networks with contrastive learning, has evolved as a pivotal tool in user-item recommendations. While promising, existing GCL methods often lack explicit modeling of hierarchical item structures, which represent item similarities across varying resolutions. Such hierarchical item structures are ubiquitous in various items (e.g., online products and local businesses), and reflect their inherent organizational properties that serve as critical signals for enhancing recommendation accuracy. In this paper, we propose Hierarchical Graph Contrastive Learning (HGCL), a novel GCL method that incorporates hierarchical item structures for user-item recommendations. First, HGCL pre-trains a GCL module using cross-layer contrastive learning to obtain user and item representations. Second, HGCL employs a representation compression and clustering method to construct a two-hierarchy user-item bipartite graph. Ultimately, HGCL fine-tunes user and item representations by learning on the hierarchical graph, and then provides recommendations based on user-item interaction scores. Experiments on three widely adopted benchmark datasets ranging from 70K to 382K nodes confirm the superior performance of HGCL over existing baseline models, highlighting the contribution of hierarchical item structures in enhancing GCL methods for recommendation tasks.

Figures (5)

• Figure 1: The framework of HGCL. Initially, HGCL generates user and item representations using XSimGCL yu2023xsimgcl, which comprises graph convolution operators with cross-layer contrastive learning (Sec. \ref{['pretraining']}). Afterwards, the learned high-dimensional item representations are projected into a two-dimensional latent space using the t-SNE algorithm, forming distinct clusters of item nodes (Sec. \ref{['clustering']}). Building upon these item clusters, HGCL constructs user-clustered item graphs with clustered items (Sec. \ref{['finetuning']}). We then jointly optimize representations of users, items, and clustered items under recommendation and contrastive losses. Finally, HGCL leverages optimized representations of users, items, and clustered items to infer personalized item rankings for each user.
• Figure 2: Learning curves of different models on the three datasets.
• Figure 3: Recall@20 scores from HGCL with different values of hyperparameters $\rho$ and $\theta$.
• Figure 4: Recall@20 and NDCG@20 scores from HGCL with different values of perplexity in t-SNE.
• Figure 5: The connecting strength of positive and negative user-item pairs after pre-training ($a$, $c$) and fine-tuning ($b$, $d$) on the Yelp2018 dataset.