Graph contrastive learning and Its applications in recommendation systems
- Publication Type:
- Thesis
- Issue Date:
- 2024
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
Graph Contrastive Learning (GCL) has emerged as a powerful tool for unsupervised graph representation learning, attracting significant attention across various applications. Its success depends on obtaining high-quality contrasting samples through graph augmentations. However, current augmentation strategies face several limitations, such as introducing noise that degrades downstream model performance, lacking flexibility for different datasets with various characteristics, and being unable to process non-embedding node features like text. These limitations hinder the full potential of GCL in practical applications. Moreover, implementing GCL in diverse application scenarios, particularly recommendation systems, is crucial for realizing its practical value. \acrfull{rs} domains are especially suitable for GCL to perform because it can generate contrasting samples that provide self-supervised training signals, addressing the lack of related information in real-world applications caused by various factors like privacy concerns . Despite its potential, the use of GCL in recommendation systems remains underexplored, and there is a need to explore its full potential in this domain.
To address the problems above, this research proposes advanced graph augmentation strategies, incorporating counterfactual mechanisms and the capabilities of \acrfull{llm}, to overcome the limitations of existing methods. By integrating counterfactual mechanisms, the proposed strategies aim to mitigate the noise introduced during graph augmentations and achieve flexibility when facing different graph data, thereby improving the performance on downstream graph learning tasks. Additionally, the utilization of LLM capabilities enables the processing of non-embedding node features like text, enhancing the flexibility of the augmentation strategies for graph data with multimodality like text features. Furthermore, this study introduces the concept of hyper meta-path to construct contrasting samples (\ie hyper meta-graphs) for GCL for multi-behavior recommendations, providing insights into creating effective contrasting samples in this specific context, which is a pioneering research work in the domain of GCL in RS and inspires the future works in the literature. This study also investigates specific training paradigms, finding that GCL pre-training and prompt-tuning can better utilize GCL’s capabilities in recommendations. By exploring these training paradigms, the research aims to provide practical guidance on how to effectively leverage GCL in recommendation systems.
In summary, the findings of this study contribute to the advancement of graph augmentation strategies for GCL and demonstrate the applicability of GCL in enhancing RS.
Please use this identifier to cite or link to this item: