用于企业风险预测的LLM增强多模态图学习框架 | 香港中文大学（深圳）经管学院

标题	用于企业风险预测的LLM增强多模态图学习框架
日期和时间	2025年12月4日（周四） 10:30 - 12:00
地点	综合教学楼D604会议室
主讲人	张东成香港中文大学商学院
摘要	Enterprise risk prediction is critical for informed investment decision-making, yet it remains challenging because it requires modeling multimodal information (e.g., structured financial indicators and unstructured text) as well as the complex interrelationships among companies and institutions. Existing methods often process these modalities in isolation or struggle to capture implicit and higher-order dependencies. To address these limitations, we propose a unified framework that integrates the strengths of large language models (LLMs) and graph learning for enterprise risk prediction using multimodal data sources. Specifically, we design a multi-stage Chain-of-Thought prompting strategy that enables LLMs to generate risk-aware textual summaries from each company’s business descriptions, financial indicators, and investor profiles. These semantic representations are fused with other structured features through a multimodal encoder. To model inter-company dependencies, we construct graphs using both investment relationships and LLM-generated summaries. Given the subtle and implicit nature of these dependencies, we introduce an adjacency augmentation mechanism that captures meaningful high-order relations and supports efficient information propagation while mitigating the over-smoothing issue in graph neural networks. Comprehensive experiments on real-world datasets spanning multiple years and markets demonstrate the superiority of our method. Overall, the proposed approach provides a novel and general LLM-enhanced multimodal graph learning framework for enterprise risk prediction.
主讲人简介	张东成是香港中文大学商学院决策、运营与科技学系的助理教授。在加入香港中文大学之前，他曾任埃默里大学戈伊苏埃塔商学院的博士后研究员。他于清华大学获得管理科学与工程博士学位、工程学学士学位及管理学学士学位。他的研究致力于开发及应用机器学习算法、统计方法与分析模型，以提升数字营销及管理信息系统领域的决策水平。他的研究兴趣尤其专注于为商业问题（如文本挖掘、消费者选择与金融科技等）开发具有可解释性或理论驱动的机器学习/深度学习算法。