Explainable Federated Learning-Based Multi-Objective Transportation Model with Dynamic Risk Assessment for Smart Logistics Networks

Abstract

Telemetry data from cross-carrier smart logistics networks form an endless flow of commercially proprietary information which, in most countries, cannot be aggregated in one place without significant legal challenges. At the same time, having a global perspective which would help identify potential delays and optimize assets in the network is critical to operation. In order to solve this dilemma, we propose a solution which keeps the telemetry data on edge devices and trains a distributed logistic regression classifier with respect to the delay risk at the same time. Our architecture consists of three parts which are typically considered separately: federated learning scheme training on a group of ten truck-level clients without any record exchange, SHAP-based post-hoc explanation module calculating explanations locally on each edge node, and a multi-objective optimization procedure balancing asset utilization against waiting time and delay risk. We treat delay risk as a dynamic value that changes over time depending on streaming traffic, weather conditions and vehicle location. The whole approach is based on publicly available Smart Logistics Supply Chain dataset where Logistics Delay Reason field is used as ground truth in order to verify whether the explanations of local models give the correct delay reason. As a result of simulation implemented in PyTorch and Flower frameworks, our federated classifier shows only two percent lower accuracy comparing to the centralized baseline but explains correctly the ground truth reason for the delay in the vast majority of cases.