SCDF-Net: Research on a Multi-Source Feature Fusion Binary Classification Prediction Network Algorithm for Supply Chain Delay Delivery Risk

Abstract

This paper addresses the problem of identifying delayed delivery risks in order fulfillment within a smart supply chain environment. It proposes a deep learning model, SCDF-Net, based on multi-source feature fusion, for unified modeling and risk assessment of complex business data. This method constructs a hierarchical feature encoding and representation learning framework around multi-dimensional heterogeneous data such as order attributes, customer information, product characteristics, and transportation and regional information. By uniformly mapping and fusing features from different sources, it effectively models the interaction relationships of multiple factors in the supply chain. During feature processing, embedded representation and structured encoding mechanisms are introduced to enhance the expressive power of categorical and continuous variables, and a multi-channel fusion strategy is combined to improve the overall feature utilization efficiency. In the risk prediction stage, the fused global representation is used to complete binary classification, thereby identifying delayed delivery risks. Experiments are conducted to compare the model's performance with various other methods. The results show that the proposed method outperforms other methods on evaluation metrics, more accurately characterizing the potential risk features in the order fulfillment process, and demonstrating strong discriminative ability and modeling effectiveness.