Abstract: With a continuous improvement in information technology, the amount of data in A-share listed companies is becoming increasingly large. How to mine and obtain information of stock market risk status, which has practical guidance value in trend prediction, from their internal and external heterogeneous data, is a significant research problem in the field of financial risk management in China. It is also one of the primary challenges in the practical oversight of financial markets. In the big data environment, massive data often lead to the “curse of dimensionality” issue in data analysis, making it a challenge for traditional single-layer or bipartite networks to efficiently represent and analyze big data of listed companies, thus making it difficult to provide timely and effective warning of stock market risks. Knowledge graphs offer innovative technical support for stock market risk contagion and prediction based on big data mining of listed companies by modeling and visualizing entities and relationships in the objective world. For A-share listed companies, the knowledge graph facilitates the visualization of both internal employment relationships, such as those between directors and executives, and various external relationships, including holdings and borrowings. Simultaneously, knowledge graph supports graph machine learning algorithms, which have exhibited notable efficacy in extracting implicit association information within enterprises and simulating risk contagion predictions.
Consequently, this paper delves into the risk contagion problem of listed companies from the perspective of knowledge graphs. Based on big data from A-share listed companies in China, it deeply analyzes the multi-layer network relationships between listed companies, constructs a knowledge graph of listed company relationships, and proposes a risk random walk model based on the personalized PageRank algorithm to numerically simulate the risk contagion process. Firstly, by employing crawler technology to obtain multi-dimensional association data of listed companies, knowledge acquisition and integration are achieved through entity disambiguation and unified processing. A top-down approach is adopted to construct the association knowledge graph of A-share listed companies. Secondly, employing the basic principles of graph theory, the correlation graph is transformed into a risk contagion graph applicable for numerical iteration simulation and prediction of risk contagion processes. Then, the personalized PageRank risk random walk model is introduced into the risk graph, proposing a risk contagion simulation model based on the personalized PageRank algorithm. The risk contagion path and the PR values of each node in the knowledge graph are obtained when they reach a steady state, identifying potential infected individuals of risk events. This enables efficient visual simulation and prediction of the contagion process of sudden risk events. Finally, the effectiveness of the risk contagion simulation method proposed in this paper is analyzed and verified using the example of the sudden risk event “ST*Longquan interest change leading to no actual controller”.
The knowledge graph constructed in this paper contains approximately 150,000 nodes and 180,000 relationships, supporting multiple functions such as visual queries, potential relationship mining, intelligent reasoning, and risk contagion simulation. From the perspective of artificial intelligence, it offers novel research perspectives and methodologies for simulating the intricate process of stock market risk contagion and efficient risk warning functions. This can provide valuable insights for related studies, including intelligent supervision of financial market risks, contributing to the advancement of intelligent monitoring, early warning, and prevention of financial risks.However, further improvements are still needed in this paper. First,the construction of a dynamic knowledge graph from a time-varying perspective has not been addressed. Second,the research on risk contagion simulation is only based on key representative shareholding relationships. In future research, deep neural network algorithms will be used to synthesize multiple associated relationships of enterprises into unified and computable risk contagion relationships, and study the simulation method of financial risk contagion based on the integration of various associated relationships. Third, the study has failed to effectively use previous samples for model training, and its risk contagion prediction accuracy can be further improved.

Key words: listed companies, knowledge graph, risk contagion simulation, personalized PageRank algorithm

摘要： 以我国A股上市公司大数据为基础,深入分析上市公司之间的多层网络关联关系,构建上市公司关联知识图谱,进而提出基于个性化PageRank算法的风险随机游走模型,对风险传染过程进行模拟。首先,运用爬虫技术获取上市公司的多维度关联数据,进而通过实体消歧和实体统一处理实现知识的获取和融合,构建A股上市公司的关联知识图谱;其次,运用图论基本原理将关联图谱转化为风险传染图谱,并将个性化PageRank风险随机游走模型引入到风险图谱中,对突发风险事件的传染过程进行高效的可视模拟和预测。本文所构建的知识图谱包含约15万个节点、18万条关系,支持可视化查询、智能化推理和风险传染模拟多重功能,从人工智能视角为金融风险传染这一复杂过程的模拟计算和高效预警提供了新的研究思路和方法,可为金融风险智能监管等研究提供有益参考。

关键词: 上市公司, 知识图谱, 风险传染模拟, 个性化PageRank