Interpretable Privacy-Preserving Collaborative Deep Learning for Taxi Trip Duration Forecasting

Abstract

Traffic data are obtained from various distributed sources such as infrastructure and vehicle sensors developed by various organisations, and often cannot be processed together because of data privacy regulations. Thus, distributed machine learning methods are required to process the data without sharing them. Federated learning allows the processing of data distributed by transmitting only the parameters without sharing the real data. The federated learning architecture is based mainly on deep learning, which is often more accurate than other machine learning approaches. However, deep-learning-based models are black-box models, and should be explained to increase trust in the system for both users and developers. Despite the fact that various explainability methods have been proposed, the solutions for explainable federated models are insufficient.

In this study, we used a federated deep learning model to predict a taxi trip duration within Brunswick region. We showed situations for which federated learning improves the prediction quality, allowing for an accuracy comparable to that obtained on the complete dataset. Moreover, we investigated how the amount of transmitted information in federated learning can be optimised while maintaining the same accuracy. Finally, we propose how the federated deep learning model can be interpreted using explainability methods without transmitting raw data and compare the results of various explainability approaches.