Federated incremental learning facilitates decentralized and continuous model updates across multiple clients, presenting a promising framework for big data analytics in distributed environments. However, the presence of poisoned or malicious data introduces significant challenges, including compromised model performance and system reliability. To tackle these issues, this paper proposes an efficient and resource-aware machine unlearning method tailored for federated incremental learning. The approach utilizes a membership inference attack mechanism to accurately identify poisoned data based on prediction confidence levels. Once detected, a targeted forgetting mechanism is applied, leveraging fine-tuning techniques to erase the influence of the poisoned data while preserving the model’s incremental learning capabilities. By aligning the distributions of poisoned data with third-party datasets, the method achieves reliable unlearning without introducing excessive computational overhead. Extensive experiments conducted on diverse datasets validate the method’s effectiveness, demonstrating a significant reduction in forgetting time (up to 21.05× speedup compared with baseline approaches) while maintaining robust model performance in incremental learning tasks. This work offers a scalable and efficient solution to the data forgetting problem, advancing the reliability and practicality of federated incremental learning in distributed and resource-constrained scenarios.

Urban taxi demand prediction faces a critical resolution paradox: high-resolution forecasts enable operational agility but suffer from extreme sparsity-induced volatility, while low-resolution predictions sacrifice responsiveness for stability. We present a Scalable SpatioTemporal Zero-Inflated Poisson Graph Neural Network (SSTZIP-GNN), that resolves this paradox through three innovations: (1) Zero-Inflated Poisson (ZIP) integration that explicitly models structural zeros in sparse demand distributions, distinguishing genuine low-demand periods from data artifacts; (2) Adaptive spatiotemporal learning that dynamically adjusts kernel dilation factors and graph diffusion rates across temporal resolutions using Diffusion Graph Convolutional Networks (DGCNs) and Temporal Convolutional Networks (TCNs); (3) Multimodal feature fusion incorporating real-time crowd-sourced mobility data, socioeconomic indicators, and Global Position System (GPS) trajectories for enhanced robustness under variable urban conditions. Extensive evaluation on 130 million real-world mobility records demonstrates superior performance, achieving 34.8% Mean Absolute Error (MAE) reduction over state-of-the-art baselines. The model reduces computational costs by 46.3% compared to ensemble approaches while maintaining high accuracy across resolutions, delivering 33.4%−53.3% Root Mean Square Error (RMSE) reduction across different prediction resolution scenarios. This unified framework enables cities to implement demand-responsive fleet management, dynamic pricing, and sustainable mobility planning across diverse urban landscapes.