Fusion Self-supervised Learning for Recommendation

TL;DR

This work tackles data sparsity in recommender systems by rethinking graph contrastive learning (GCL). It introduces High-order Fusion Graph Contrastive Learning (HFGCL), which constructs high-order contrastive views by excluding low-order information and fuses multiple self-supervised signals via a fusion CL loss, avoiding costly data augmentations. Empirically, HFGCL achieves state-of-the-art performance and faster training across three public datasets, outperforming augmentation-based CL methods and other baselines, especially under data sparsity. The results suggest that high-order views and signal fusion offer a practical, scalable path for integrating self-supervised signals into recommender systems, with potential for further insights into CL interpretability.

Abstract

Recommender systems are widely deployed in various web environments, and self-supervised learning (SSL) has recently attracted significant attention in this field. Contrastive learning (CL) stands out as a major SSL paradigm due to its robust ability to generate self-supervised signals. Mainstream graph contrastive learning (GCL)-based methods typically implement CL by creating contrastive views through various data augmentation techniques. Despite these methods are effective, we argue that there still exist several challenges. i) Data augmentation ($e.g.,$ discarding edges or adding noise) necessitates additional graph convolution (GCN) or modeling operations, which are highly time-consuming and potentially harm the embedding quality. ii) Existing CL-based methods use traditional CL objectives to capture self-supervised signals. However, few studies have explored obtaining CL objectives from more perspectives and have attempted to fuse the varying signals from these CL objectives to enhance recommendation performance. To overcome these challenges, we propose a Fusion Self-supervised Learning framework for recommendation. Specifically, instead of facilitating data augmentations, we use high-order information from GCN process to create contrastive views. Additionally, to integrate self-supervised signals from various CL objectives, we propose an advanced CL objective. By ensuring that positive pairs are distanced from negative samples derived from both contrastive views, we effectively fuse self-supervised signals from distinct CL objectives, thereby enhancing the mutual information between positive pairs. Experimental results on three public datasets demonstrate the superior recommendation performance and efficiency of HFGCL compared to the state-of-the-art baselines.

Figures (6)

• Figure 1: GCL-based methods using data augmentation. Top: Graph augmentation; Bottom: Feature augmentation.
• Figure 2: Framework of HFGCL. Left: User-item interaction. Center: High-order contrastive views generation process. In the figure, let $h = 2$ and $L = 3$. After aggregating high-order embeddings, user $u_2$ has a representation more similar to the positive item $i_2$, with only $u_4$ missing. This indicates a close relationship between $u_2$ and $i_2$ in the high-order embedding space. Right: Dual optimization objectives. During the model training phase, two optimization objectives are included: the primary and the contrastive optimization objective, refer to Fig. \ref{['fig:negative_sample']} for the meaning.
• Figure 3: HFGCL Optimization Process. This process optimizes the primary objective, while fusing user, item, and user-item self-supervised signals from diverse contrastive views via a fusion CL objective.
• Figure 4: Total training costs comparison of HFGCL and other GCL-based methods on three datasets.
• Figure 5: Performance comparison w.r.t. NDCG@20 for different user groups sparsity levels on three datasets.