Abstract
Personalized career recommendations for university students require models that can capture evolving preferences and support long-term career-development decision-making. However, conventional matching-based approaches often struggle to effectively model temporal dynamics and integrate heterogeneous user information within a unified framework. To address these limitations, this study proposes a hierarchical temporal reinforcement learning framework (HRA-TS) for personalized career recommendation. Specifically, the framework employs a hybrid long short-term memory–graph attention network encoder to jointly capture users’ evolving behavioral evolution and stable personal attributes. A multimodal reward mechanism is introduced to incorporate matching relevance, developmental potential, and behavioral feasibility, thereby guiding policy learning toward more realistic career trajectories. To further enhance training efficiency and decision stability, a curriculum learning strategy is adopted to progressively expand the recommendation space. In addition, federated learning is incorporated to enable privacy-preserving training across distributed clients. Experimental results demonstrate the effectiveness of the proposed framework. Compared with the deep Q-network baseline, HRA-TS improves long-term job matching by 6.02% and cold-start click-through rate by 6.09%. Under the federated learning setting, performance degradation remains below 2.9%, indicating that privacy preservation can be achieved with only limited accuracy loss. These findings suggest that HRA-TS provides an effective and practical solution for personalized career recommendation in dynamic and privacy-sensitive environments.