A1184
Title: Privacy-aware delivery classification via federated learning
Authors: Hang Qi - BNU-HKBU United International College (China) [presenting]
Abstract: Federated learning (FL) provides a privacy-preserving framework for collaborative classification, especially in supply chain settings where vendors are reluctant to share sensitive data. The aim is to predict delivery anomalies, such as late shipments and abnormal logistics events, under a FL framework. Personalized and adaptive decomposition for federated learning (PAD-FL) is proposed, focusing on classification performance, adaptability, and privacy-aware delivery. Logistic regression is compared as a centralized baseline with MLP and LSTM models under both federated averaging (FedAvg) and PAD-FL. To simulate real-world heterogeneity, non-IID client partitions are constructed based on customer segments, regions, and product categories. While FedAvg provides a simple aggregation strategy, it suffers from performance degradation in non-IID scenarios due to model drift. In contrast, PAD-FL effectively mitigates this by decoupling the model into shared global and personalized local components, allowing better adaptation to client-specific distributions. Using a real-world supply chain dataset with diverse customer, product, and regional attributes, a node selection mechanism is further proposed based on local statistical indicators skewness, kurtosis, and covariance to improve aggregation reliability. All methods are evaluated in terms of classification accuracy and communication cost, offering insights into scalable and personalized FL systems for supply chain analytics.
