Hyperbolic Contrastive Learning with Model-augmentation for Knowledge-aware Recommendation

TL;DR

This work tackles knowledge-aware recommendation by addressing two core challenges: capturing hierarchical structure in user–item and knowledge graphs, and avoiding preference shifts caused by graph perturbations. It introduces HCMKR, which employs Lorentzian knowledge aggregation within a hyperbolic space to encode users and items, and augments training with three model-level contrastive strategies (Dropout, cross-layer outputs, and model pruning) to provide robust supervisory signals without altering the input graph. The method jointly optimizes a hyperbolic recommendation loss and a contrastive loss, achieving up to 11% improvements over strong baselines while maintaining competitive training efficiency. Empirical results on Yelp2018, Amazon-Book, and MovieLens-20M demonstrate the effectiveness of the hyperbolic, Lorentzian framework and model-level augmentations for knowledge-aware recommendation, with clear ablations supporting the contributions and findings.

Abstract

Benefiting from the effectiveness of graph neural networks (GNNs) and contrastive learning, GNN-based contrastive learning has become mainstream for knowledge-aware recommendation. However, most existing contrastive learning-based methods have difficulties in effectively capturing the underlying hierarchical structure within user-item bipartite graphs and knowledge graphs. Moreover, they commonly generate positive samples for contrastive learning by perturbing the graph structure, which may lead to a shift in user preference learning. To overcome these limitations, we propose hyperbolic contrastive learning with model-augmentation for knowledge-aware recommendation. To capture the intrinsic hierarchical graph structures, we first design a novel Lorentzian knowledge aggregation mechanism, which enables more effective representations of users and items. Then, we propose three model-level augmentation techniques to assist Hyperbolic contrastive learning. Different from the classical structure-level augmentation (e.g., edge dropping), the proposed model-augmentations can avoid preference shifts between the augmented positive pair. Finally, we conduct extensive experiments to demonstrate the superiority (maximum improvement of $11.03\%$) of proposed methods over existing baselines.

Figures (6)

• Figure 1: Degree distributions of two public datasets.
• Figure 2: Illustrations of the model-level augmentations and preference shift. (a) An abstract process of representation learning in the GNN-based knowledge-aware recommender systems, where $f(\cdot)$ is the abstract encoding function. We mainly focus on perturbing the structure of $f(\cdot)$ to generate augmented samples. (b) Augmentations via cross-layer outputs. (c) Augmentations via Dropout. (d) Augmentations via model-pruning. (e) The structural perturbation leads to the loss of the user preference for electronic products.
• Figure 3: Visualizing the distribution of embeddings (Yelp2018).
• Figure 4: Visualizing the distribution of embeddings (Amazon-Book).
• Figure 5: Ablation experiments of the proposed methods.