Federated Learning assisted framework to periodically identify user communities in urban space

Abstract

Identifying individuals with similar behaviors and mobility patterns has become essential to improving the functioning of urban services. However, since these patterns can vary over time, such identification needs to be done periodically. Furthermore, once mobility data expresses the routine of individuals, privacy must be guaranteed. In this work, we propose a framework for periodically detecting and grouping individuals with behavioral similarities into communities. To accomplish this, we built an autoencoder model to extract spatio-temporal mobility features from raw user data at periodic intervals. We used Federated Learning (FL) as a training approach to preserve privacy and alleviate time-consuming training and communication costs. To determine the number of communities without risking an arbitrary number, we compared the choices of two probabilistic methods, the Akaike Information Criterion (AIC) and the Bayesian Information Criterion (BIC). Since the communities are updated periodically, we also analyzed the impact of aged samples on the proposed framework. Finally, we compared the performance of our FL-based solution to a centralized training approach. We analyzed similarity and dissimilarity metrics on mobility samples and the contact time of individuals in three different scenarios. Our results indicate that AIC outperforms BIC when choosing the number of communities, although both satisfy the evaluation metrics. We also found that using older samples benefits more complex spatio-temporal scenarios. Finally, no significant losses were detected when compared to a centralized training approach, reinforcing the advantages of using the FL-based method.