Network-State-Intelligent Learning Rate Optimization for Federated Streaming Models in Smart Transportation

Data-driven traffic modeling in Intelligent Transportation Systems (ITS) often learns traffic trends offline from massive historical databases rather than continuously learning in near-real-time as new data are collected, which can help capture the impacts of irregular incidents with low periodic representation (nonrecurrent events (NREs)). Traditional approaches face challenges when predicting the impacts of NREs due to their low frequency in the training data, often leading to their removal or smoothing during preprocessing. Online federated learning (FL), a privacy-preserving and distributed approach, offers a promising alternative. It enables models to continuously learn and update their representations using newly collected data from online data streams. It presents the opportunity to improve the model’s ability to dynamically adapt to stochastic network changes. In conjunction with online FL, variable learning rate strategies can enhance the model’s ability to learn stochastic spatial-temporal trends reliably, including NREs. This paper designs and analyzes the impact of two novel traffic-aware adaptive learning rate strategies, one time-dependent strategy and another dependent on the network state and existence of NREs, for improving the traffic representations learned through online FL. We design a novel heuristic that operates at the client level and dynamically adapts the local learning rate in response to multiple parameters. We conducted extensive experiments using three datasets containing different traffic variables, and our results demonstrated that our learning rate strategies reduced the online inference error by about 15.77% and enhanced the FL models’ resilience to stochastic traffic scenarios.