KPSR: Knowledge-Enhanced Hierarchical Model for Personalized Session-Based Recommendation

Industrial Internet of Things environments continuously produce sequential operational data, such as sensor streams, maintenance logs, and service records, making next-action recommendation increasingly important for smart manufacturing and predictive maintenance. Personalized session-based recommendation (PSBR) provides a natural framework for such tasks by combining short-term session intent with long-term behavioral history. However, existing PSBR methods generally neglect two key requirements of industrial applications: the structured knowledge associated with industrial entities and the unequal contribution of historical sessions to the current decision context. To address these challenges, we propose a knowledge-enhanced personalized session-based recommendation (KPSR) model. KPSR constructs a heterogeneous knowledge graph by jointly encoding historical session transitions and item attribute relations, and leverages TransR to learn knowledge-enhanced item representations. On this basis, we develop a hierarchical session encoder to model sequential preferences at both intrasession and intersession levels. Specifically, the internal session encoding layer captures fine-grained transitions within the current session, while the external session encoding layer adaptively measures the relevance of historical sessions. Experiments on four real-world datasets show that KPSR consistently outperforms state-of-the-art baselines.