POMNS: Positive-Offset Mixed Negative Sampling for Graph-Based Recommendation

Hard negative sampling has achieved great success in recent years across various deep learning tasks because of its ability to accelerate model convergence and improve performance. In recommender systems, due to the inherently sparse nature of interaction data makes it particularly challenging to mine informative and representative hard negatives. Existing methods often rely on predefined heuristic sampling strategies to search for valuable negative samples within the massive item space, and typically suffer from high computational overhead, low sampling efficiency, and unstable sample quality, significantly limiting their practical applicability in real-world recommendation scenarios. In this paper, we propose a novel and efficient negative sampling method to generate informative hard negatives through subtle perturbations of positive samples, called the Positive-Offset Mixed Negative Sampling strategy (POMNS). This yields a semantically challenging yet informative hard negative sample, while avoiding excessive computational cost and enabling the generation of high-quality training signals more efficiently. Furthermore, we extend the strategy to multi-layer graph neural networks to fully exploit the structural information. The sampling result is a hierarchical hard negative embedding that fuses high-order structural features with semantic signals for more effective representation learning. Extensive experiments on three benchmark datasets demonstrate that POMNS consistently improves recommendation performance.