SCONE: A Novel Stochastic Sampling to Generate Contrastive Views and Hard Negative Samples for Recommendation

TL;DR

SCONE tackles data sparsity and negative sampling in graph-based recommender systems by unifying contrastive view generation and hard negative sampling under score-based generative models. It combines LightGCN as the encoder with a learned score network to produce personalized augmented views and hard negatives through stochastic forward and reverse diffusion processes. The approach yields superior performance across six datasets, improves robustness to user sparsity and item long-tail distributions, and enhances representation uniformity and diversity. This work offers a practical, scalable pathway to more accurate and fair recommendations by leveraging stochastic sampling in a joint framework.

Abstract

Graph-based collaborative filtering (CF) has emerged as a promising approach in recommender systems. Despite its achievements, graph-based CF models face challenges due to data sparsity and negative sampling. In this paper, we propose a novel Stochastic sampling for i) COntrastive views and ii) hard NEgative samples (SCONE) to overcome these issues. SCONE generates dynamic augmented views and diverse hard negative samples via a unified stochastic sampling approach based on score-based generative models. Our extensive experiments on 6 benchmark datasets show that SCONE consistently outperforms state-of-the-art baselines. SCONE shows efficacy in addressing user sparsity and item popularity issues, while enhancing performance for both cold-start users and long-tail items. Furthermore, our approach improves the diversity of the recommendation and the uniformity of the representations. The code is available at https://github.com/jeongwhanchoi/SCONE.

Figures (13)

• Figure 1: Our proposed user-item specific stochastic sampling method based on score-based generative models for contrastive learning and negative sampling. We generate contrastive views for CL with a stochastic process and hard negative samples with a stochastic positive injection.
• Figure 2: The overall framework of score-based generative models. SGMs corrupt the input data with increasing noise and generate a new sample from the noisy data.
• Figure 3: The overall workflow of SCONE. We generate contrastive views and hard negative samples via stochastic sampling based on score-based generative models, and then utilize them for encoder training.
• Figure 4: Performance comparison over different user groups. More results are in oAppendix \ref{['app:robustness']}.
• Figure 5: Performance comparison over different item groups. More results are in Appendix \ref{['app:robustness']}.