Federated LSTM-based deep learning model for privacy-preserving predictions of human trajectories across multiple data providers

The problem of data-driven location prediction of individual users is based on an effective mining of travel behaviours and motion patterns. In this sense, a central aspect refers to the use of a relevant amount of historical mobility data, required for successfully training machine learning models, especially when involving artificial neural networks. However, such data are sensitive in nature, therefore not easily available and always subjected to privacy-related restrictions on their public share. With the purpose of merging information from different providers without directly sharing geo-private data, we hereby assess the feasibility of decentralized training over multiple data sources, leveraging different unmergeable trajectory datasets stored in separate servers. In particular, we integrate a long short-term memory (LSTM) recurrent neural network framework for location prediction into a federated learning environment, whereby local workers compute the network operations on their data share, and the learning results are progressively synchronized with a parameter server. Variants of federated algorithms are evaluated and compared to separate independent training processes (lower benchmark) and an ideal, but in fact not allowed, centralized training (upper benchmark). By leveraging real-world datasets of sparse non-repetitive mobility traces, our experiments aim to disclose insights on federated learning strategies for advanced trajectory analytic tasks, paving the way to decentralized applications involving multiple geo-private data sources.