Socially-Aware Self-Supervised Tri-Training for Recommendation (2106.03569v4)

Abstract: Self-supervised learning (SSL), which can automatically generate ground-truth samples from raw data, holds vast potential to improve recommender systems. Most existing SSL-based methods perturb the raw data graph with uniform node/edge dropout to generate new data views and then conduct the self-discrimination based contrastive learning over different views to learn generalizable representations. Under this scheme, only a bijective mapping is built between nodes in two different views, which means that the self-supervision signals from other nodes are being neglected. Due to the widely observed homophily in recommender systems, we argue that the supervisory signals from other nodes are also highly likely to benefit the representation learning for recommendation. To capture these signals, a general socially-aware SSL framework that integrates tri-training is proposed in this paper. Technically, our framework first augments the user data views with the user social information. And then under the regime of tri-training for multi-view encoding, the framework builds three graph encoders (one for recommendation) upon the augmented views and iteratively improves each encoder with self-supervision signals from other users, generated by the other two encoders. Since the tri-training operates on the augmented views of the same data sources for self-supervision signals, we name it self-supervised tri-training. Extensive experiments on multiple real-world datasets consistently validate the effectiveness of the self-supervised tri-training framework for improving recommendation. The code is released at https://github.com/Coder-Yu/QRec.

• The paper introduces SEPT, a framework that integrates self-supervised learning with tri-training to leverage social network homophily for enhanced recommendations.
• It employs a novel neighbor-discrimination-based contrastive learning method to refine multi-view graph representations using signals from adjacent nodes.
• Empirical validation on Last.fm, Douban-Book, and Yelp datasets demonstrates statistically significant performance gains, particularly in sparse data scenarios.

Review of "Socially-Aware Self-Supervised Tri-Training for Recommendation"

The paper presents a novel approach to enhancing recommender systems using a framework called Socially-Aware Self-Supervised Tri-Training (SEPT). This framework leverages self-supervised learning (SSL) to harness supervisory signals from data without labels. At the core of the proposed method is the integration of tri-training and SSL, designed to capture homophily within social networks for improved recommendation performance.

Core Contributions

1. Framework Design: The true innovation of SEPT lies in its ability to synergize SSL with tri-training to fully exploit social information for recommender systems. By constructing three distinct graph encoders for multi-view encoding, SEPT enriches user-item interaction data with social context, utilizing auxiliary views derived from user relations.
2. Contrastive Learning: The authors promote a novel neighbor-discrimination-based contrastive learning methodology. This strategy contrasts with conventional self-discriminative methods by utilizing supervisory signals not only from the node itself but also from neighboring nodes, aligning with the homophily observed in social networks.
3. Empirical Validation: Using real-world datasets (Last.fm, Douban-Book, and Yelp), the performance improvements achieved by SEPT were statistically significant. On less dense datasets, the framework showed particularly pronounced benefits, suggesting its utility in scenarios where data sparsity is a challenge.

Detailed Analysis

The paper systematically breaks down the SEPT framework. Initially, the authors introduce the concept of leveraging multi-view data sources. They enhance the traditional user-item interaction graph with two additional views constructed from social relations by identifying triadic structures. This approach is particularly insightful as it implicitly profiles user interests derived from their social interactions.

The tri-training component initiates with three encoders operating on these diverse views. By combining user-item interactions with user-to-user recommendations, the framework iteratively refines each encoder’s representations, extracting labels generated by alternative encoders. This dynamic interaction enhances the adaptability and strength of recommendations.

A major highlight is the framework's adaptability to multiple AI models, showcasing flexibility and potential applicability across various domains beyond recommendation systems. Specifically, the use of LightGCN as the base structure highlights the simplicity and efficiency of the proposed system, providing a robust baseline for further exploration.

Implications and Future Directions

The SEPT framework's implications extend to both theoretical and practical dimensions. The method's novel integration of SSL with tri-training paves the way for more nuanced and contextually aware recommendation systems. Practically, its ability to exploit homophily offers a path toward more personalized and accurate recommendations. The promising results suggest exciting prospects for deploying such systems across industries relying heavily on recommendation engines, such as e-commerce and entertainment.

Future exploration could focus on extending the framework to include item-level self-supervision or incorporating multimodal data. These enhancements might uncover even deeper insights into user behavior, further broadening the scope and applicability of SEPT.

In summary, the paper delivers a comprehensive contribution to the field of recommendation systems through its inventive use of socially-aware SSL. By unfolding the potential of multi-view co-training, it sets a substantial precedent for future innovations in AI-driven personalization.