Graph Neural Networks for Default Risk Prediction and Recommendation
- Publication Type:
- Thesis
- Issue Date:
- 2025
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
Encouraged by the incredible success of graph neural networks on a broad spectrum of practical applications, using this technology for credit risk prediction and recommendation has become ubiquitous recently. Specifically, home credit default risk prediction aims to detect clients and loan applications that have the potential risk of failing to repay a loan or meet contractual obligations. It is essential for banks and financial institutions to keep good health in management and operations. Recommendation systems endeavor to assist users in finding valuable information or potential products they might prefer from the massive amount of items, rendering the decision-making process easier. These two tasks encounter a huge challenge of graph data processing and mining, and both of them attract widespread attention from academia and industry. In this thesis, we mainly focus on the graph neural networks for default risk prediction and recommendation. More concretely, we apply graph neural networks to alleviate the data missing issue via information propagation and aggregation of similar records. Besides, multi-view graphs are designed for small data augmentation to tackle the unbalanced and skewed distribution problem. So as to the session-based recommendation, we first give a comprehensive review of the graph and sequential neural networks in this task. Then, a dual graph neural network is proposed to capture the implicit and explicit relationships among the external and internal session connections, which will be fused together for the final recommendation. Also, considering the skewed item distribution in the recommendation system, a reweighing and reembedding strategy is applied to alleviate the tail item problem in text-based or image-based recommendation. We conducted extensive experiments with regard to the default risk prediction and recommendation tasks on three and six datasets against fifteen and twenty-four baselines, respectively, to verify the effectiveness of our methods. We believe the proposed multi-view and dual graph neural networks are capable of transferring and facilitating a wide range of real-world domains, including medicine, traffic, social networks, and beyond.
Please use this identifier to cite or link to this item: