A federated meta-learning approach for interpretable, privacy-preserving, and customizable behavior analysis

Travel behavior analysis provides critical insights to enhance the intelligence of transportation systems, enabling more accurate and efficient management of mobility services. However, it requires centralizing user-sensitive data which may violate regulations and laws about data security. Even though various solutions have been proposed to train deep learning models (DNNs) via federated learning, it still faces three critical challenges in ensuring the interpretability of DNNs to unfold the black-box, bridging isolated data to train adaptive model, and harnessing the heterogeneity among users to support personalized analysis. To tackle these challenges, this paper proposes an interpretable, privacy-preserving and customizable approach to support travel behavior analysis based on federated meta-learning, called IPC-FM. Specifically, it, first, introduces an artificial neural network empowered with three kinds of utilities associated with discrete choice models to provide interpretable results. Second, it integrates federated meta-learning to train a globally meta-model via the knowledge among clients in a collaborative and privacy-preserving manner. Finally, it enables rapid model localization to support personalized analysis. Based on standard datasets, IPC-FM is evaluated against state-of-the-art methods. Results show that IPC-FM can collaborate clients with isolated and heterogeneous data to train a robust, customizable and interpretable model for travel behavior analysis.