Abstract:

During the process of enterprise investment attraction， there are multi-dimensional risks. Traditional risk assessment methods are difficult to identify these risks timely and accurately due to information distortion and complex relationships in economic behaviors. To address the above issues， a risk analysis framework integrating Large Language Model （LLM） and Graph Neural Network （GNN） was proposed. The semantic understanding capability of LLM was utilized to assist the GNN in constructing a more comprehensive and accurate dynamic heterogeneous knowledge graph of enterprises， thereby solving the information distortion problem caused by static data. On this basis， to address the shortcomings of GNN in terms of deep and semantic expression abilities， a knowledge-based semantic structure mining module was designed， and Qwen large model was combined to enhance the semantic accuracy of node representations. Furthermore， an Integrated One Graph （IOG） module was proposed to unify node classification and graph classification tasks into the prediction of “focus nodes”. Through a unified prediction mechanism， predictions for different graph structure types were achieved， thereby improving the model’s generalization ability on different datasets significantly. The IOG-CIQAN（In One Graph with Collective Intelligence and Qwen2 Assistance Network） model constructed on the basis of this framework achieved accuracy over 87% on all of three risk analysis datasets in labor， finance， and administration compared to multiple baseline models such as Capsule Network （CapsNet）.

Key words: Graph Neural Network (GNN), Large Language Model (LLM), graph structure awareness, enterprise risk prediction, unified graph structure representation

摘要：

在企业的招商引资过程中，存在多维度的风险。传统的风险评估方法由于信息失真以及经济行为中的复杂关系，难以及时且准确地识别这些风险。为解决上述问题，提出一种将大型语言模型（LLM）与图神经网络（GNN）融合的风险分析框架。利用LLM的语义理解能力，辅助GNN构建全面、准确的动态企业异构知识图谱，从而解决静态数据引起的信息失真问题。在此基础上，针对GNN在深度和语义表达能力上的不足，设计一个基于知识的语义结构挖掘模块，并结合Qwen2大模型增强节点表示的语义精准性。此外，提出一体化图（IOG）模块将节点分类与图分类任务统一为对“关注节点”的预测。通过统一预测机制，实现对不同图结构类型的预测，从而显著提升模型在不同数据集上的泛化能力。基于该框架构建的IOG-CIQAN（In One Graph with Collective Intelligence and Qwen2 Assistance Network）模型在劳工、财务、行政这3个风险分析数据集上的准确率均超过了87%，优于胶囊网络（CapsNet）等多种基线模型。

关键词: 图神经网络, 大语言模型, 图结构感知, 企业风险预测, 图结构统一表示