A Machine Learning Framework for Enterprise Risk Prediction: Unified Feature Embedding and Lightweight Attention

This paper addresses the problem of enterprise risk identification in complex and dynamic business environments by proposing a risk prediction method based on lightweight neural networks. A unified data representation is first constructed to embed multi-source heterogeneous features of enterprises, allowing inputs to be effectively expressed in a low-dimensional space. The model structure combines grouped convolution and depthwise separable convolution to reduce parameter size and computational complexity while preserving the ability to capture key patterns. To further enhance the modeling of temporal dependencies, a lightweight attention mechanism is integrated to highlight critical time segments through weighted feature aggregation. In the output layer, fully connected layers and regularization strategies are used for prediction and generalization control. In addition, sensitivity experiments are conducted to analyze the effects of batch size, hidden dimension, convolution grouping, class imbalance ratio, and cold-start ratio on single performance metrics. The results show that different parameters and environmental conditions have significant impacts on the model’s discriminative ability, precision, and recall. Through these designs and analyses, the study verifies that lightweight neural networks can balance efficiency and accuracy in enterprise risk prediction tasks and provide a reliable reference for future research on risk modeling.