High-Order Graph Relationship-Based Sampling Strategies for Recommendation Systems

Zhiqiang Huang

Keywords: Recommendation systems, Collaborative filtering, Sampling strategies, Negative sampling, Curriculum learning

TL;DR: In this work, we introduced high-order neighbor relationship-based sampling strategies which aim to leverage graph structural information at the data level, and the SSB(Sampling Strategy Block) training framework

Abstract: In recommendation systems, collaborative filtering typically leverages user-item interaction history to encode user preferences and item characteristics into dense vectors. Most existing research focuses on model architecture, such as LightGCN, which explicitly incorporates high-order graph neighbor relationships into the embedding function, significantly enhancing model performance. However, research on data-level strategies, particularly sampling methods, remains limited. Most approaches simply treat items interacted with by users as positive samples, while all other items are considered negative samples. Negative samples are then randomly selected or sampled based on popularity. In contrast, we argue that high-order graph relationships can be exploited at the data level to generate higher-quality positive and negative samples, and can also serve as a form of data augmentation. Furthermore, we introduce a curriculum learning-based sampling strategy block (SSB) as a novel training approach. Through extensive experiments, we analyze the impact of the proposed sampling strategies and validate the effectiveness of SSB.

Submission Number: 26