Shaopeng Guan

Sequential recommendation aims to predict a user’s next interaction by modeling the sequential dependencies in historical behaviors. However, Transformer-based methods incur quadratic computational cost on long sequences and often struggle to capture behavioral patterns at different temporal scales, particularly short-term interest shifts. To address these limitations, we propose SEMamba, a sequential recommendation model based on state space models. SEMamba employs a Mamba encoder to model global dependencies and capture long-term user preferences. It further introduces a local spatial enhancement module to extract short-term interest patterns from neighboring interactions. A position-wise gating mechanism adaptively modulates the locally aggregated features before integrating them with the Mamba output. Experiments on the ML-1M, KuaiRand, and Mind datasets show that SEMamba improves Hit Rate by 1.60%, NDCG by 2.01%, and MRR by 2.23% on average over strong baselines. The results also demonstrate improved training stability and better generalization across datasets.

High mobility and non-independent and identically distributed (non-IID) data in the Internet of Vehicles (IoV) severely limit the communication efficiency and convergence of federated learning. This paper proposes an adaptive hierarchical federated learning framework that jointly improves communication efficiency, model accuracy, and fairness through downlink knowledge guidance and uplink gradient compression. The framework forms an asymmetric training loop in which multi-dimensional knowledge distillation transfers edge logits and intermediate features to guide heterogeneous vehicles toward fast feature alignment, thereby mitigating non-IID conflicts. On this basis, an adaptive gradient compression mechanism enables roadside units to measure the cosine similarity and magnitude of local updates in real time and to compute sparsification thresholds dynamically so as to reduce communication cost while preserving model accuracy. A dynamic client selection strategy that combines long-term reputation with ϵ-greedy exploration further decouples efficiency from fairness and alleviates the winner-takes-all problem. Experiments based on real traffic trajectories and vision benchmarks show that the proposed framework reduces communication overhead by 94.3% compared with standard federated averaging. It achieves 98.17% test accuracy on GTSRB and improves accuracy by up to 6.01 percentage points over the best baseline on CIFAR-100, while effectively exploiting data heterogeneity for additional performance gains.