Traffic flow prediction is fundamental to traffic management and scheduling decisions in intelligent transportation systems. Traffic flow data exhibit complex spatiotemporal dynamics, multi-scale nonlinearities, and pronounced trend and seasonal patterns. Consequently, conventional statistical methods and univariate time-series models often struggle to balance predictive accuracy with computational efficiency. To address these challenges, we propose a multi-scale dynamic spatio-temporal graph trend-seasonality network (MS-DSTSN). First, we design a lightweight adaptive dynamic graph learning module. It leverages low-rank tensor decomposition and threshold-based sparsification to reduce graph construction complexity from O ( D 3 ) to O ( D K 2 ) . This reduction enables real-time capture of time-varying spatial correlations among nodes. Second, we develop a dynamic multi-scale spatio-temporal embedding module comprising three temporal scales and two spatial scales. We then employ an attention mechanism to adaptively fuse features across these scales. Third, we integrate a trend-seasonality decomposition module to explicitly separate traffic flow sequences into trend and seasonal components. This enhancement improves the model’s ability to recognize long-term periodic variations and abrupt fluctuations. Finally, we adopt an encoder–decoder architecture featuring a local encoder based on a gated recurrent unit and a global encoder based on a Transformer. This design captures both short-term dynamics and long-range dependencies. In multi-step forecasting experiments on real-world traffic datasets, MS-DSTSN achieves superior results across various evaluation metrics. Furthermore, ablation studies validate the effectiveness of each module. By balancing predictive accuracy, interpretability, and computational efficiency, MS-DSTSN offers a practical solution for large-scale urban traffic flow forecasting.

With the development of the advanced educational platform, more and more personalized recommendation systems are used to increase the efficiency of learning and make the dissemination of the educational resources easier. However, educational recommendation tasks involve highly sensitive student data. The current centralized recommendation methods have limitations in learning diverse learner preferences across institutions and are susceptible to privacy violations.The existing federated recommendation methods mainly focus on the recommendation effectiveness or privacy protection. However, few works consider privacy protection, sequential learner behavior modeling and personalized federated optimization in a unified framework. To tackle these challenges, in this work we propose a privacy preserving personalized federated recommendation system for educational data.The study propose a framework that incorporates differential privacy perturbation, personalized federated aggregation, sequential interest modeling, and context-aware ranking into the optimization objective. Local learner representation is learned with differential privacy constraints. Besides, sequential behavior encoding and attention-driven modeling are used to understand the dynamic interests of learners and the continuous relation of behaviors. Besides, we present a local-global representation fusion strategy to address the dilemma between learner customization and global generalization for non-independent and identically distributed (non-IID) federated scenarios.Extensive experiments under the IID, non-IID and privacy-preserving federated frameworks are performed on the MovieLens-1 M and Amazon Beauty datasets. Our results show that the proposed Differential Privacy Federated Averaging (DP-FedAvg) framework can effectively address the performance degradation problem caused by the diverse client distributions, and maintain the recommendation accuracy under different privacy budgets. In the MovieLens-1 M dataset, the DP-FedAvg achieves an average HR@10 of 0.710 and an average Normalized Discounted Cumulative Gain (NDCG@10) of 0.681, outperforming the usual non-IID federated benchmarks. The privacy-utility evaluation further shows a nice coexistence of the privacy protection and the usefulness of recommendations. We propose a framework that provides a unified and repeatable methodology for privacy-aware federated recommendations in smart educational environments in a decentralized manner.