ABSTRACT:
Extracting effective features from dynamic networks underpins the development of network-based artificial intelligence (AI) methods and decision support systems. Despite existing methods for constructing network features, a notable gap exists in addressing the propagation influence of direct and indirect neighbors. Given the sparse and interacting nature of neighbor information as well as the requirements of interpretability, modeling neighbor influence presents an essential yet challenging research problem. To tackle this challenge, we propose a novel Time-vAriant Graph Contrastive Learning method (TAGOL). TAGOL seeks to improve both the effectiveness and interpretability of constructing features related to the propagation influence by explicitly modeling sparse and interacting neighbor information in time-variant networks. We perform a comprehensive evaluation of the proposed method through two case studies: credit risk prediction and financial distress prediction. Experimental results demonstrate the efficacy of TAGOL and shed light on the varied influences of the joint propagation of interacting neighbor information on financial risk prediction. The proposed TAGOL and experimental findings offer generalizable methodological and theoretical insights, which can contribute to a broader spectrum of network-related research endeavors, such as short video recommendation systems and transit flow prediction.
Key words and phrases: Propagation influence, social networks, network neighbors, financial risk prediction, graph neural networks, time-variant network, interpretability, machine learning, AI, artificial intelligence, decision support, predictive modeling