The trend of conglomerate and globalization in enterprise development not only complicates the enterprise financial network, but also makes capital flow, equity relationship and resource allocation efficiency gradually become the key factors affecting the comprehensive performance of enterprises. In this paper, from the financial and non-financial perspectives, we initially design 36 corporate financial risk early warning indicators under 9 different dimensions. Then it proposes a method to mine higher-order topological features in relational networks to improve the prediction accuracy of corporate financial risk dilemmas. The method constructs higher-order topological structure features by constructing a heterogeneous graph representation learning model, and identifies the higher-order topological structures that occur frequently and have an important impact on corporate financial risk. Subsequently, the weights and thresholds of the BP neural network are optimized by using the Tennessee whisker search method, so as to establish the financial risk early warning system based on BAS-BP neural network and complete the construction of the enterprise financial risk early warning model. At the same time, the 36 selected three-level indicators were tested for normality and non-parametric test, and finally the enterprise financial risk early warning indicators with 8 common factors as the main components and 27 three-level indicators were established. The proposed model is used to predict the financial fraud risk of enterprise P during a total of five years from 2017 to 2021, and its prediction results have a coefficient of determination greater than 0.900, and the root-mean-square error is less than 0.200, which is much better than similar modeling methods.
