Efficient Model-Agnostic Continual Learning for Next POI Recommendation

TL;DR

This work addresses continual next POI recommendation, where models must adapt to evolving user check-ins without full retraining. It introduces GIRAM, a model-agnostic framework that combines an interest memory, context-aware key encoding, generative key-based retrieval, and adaptive update and fusion to balance historical and recent interests. Across three real-world datasets and multiple backbones, GIRAM consistently improves accuracy (Acc@5/10/20, MRR) while maintaining low update time and memory usage, outperforming static baselines and often rivaling retraining. The results demonstrate the practicality of continual adaptation for NPR systems and highlight the value of diverse, context-rich memory retrieval coupled with adaptive memory updates for real-world deployment.

Abstract

Next point-of-interest (POI) recommendation improves personalized location-based services by predicting users' next destinations based on their historical check-ins. However, most existing methods rely on static datasets and fixed models, limiting their ability to adapt to changes in user behavior over time. To address this limitation, we explore a novel task termed continual next POI recommendation, where models dynamically adapt to evolving user interests through continual updates. This task is particularly challenging, as it requires capturing shifting user behaviors while retaining previously learned knowledge. Moreover, it is essential to ensure efficiency in update time and memory usage for real-world deployment. To this end, we propose GIRAM (Generative Key-based Interest Retrieval and Adaptive Modeling), an efficient, model-agnostic framework that integrates context-aware sustained interests with recent interests. GIRAM comprises four components: (1) an interest memory to preserve historical preferences; (2) a context-aware key encoding module for unified interest key representation; (3) a generative key-based retrieval module to identify diverse and relevant sustained interests; and (4) an adaptive interest update and fusion module to update the interest memory and balance sustained and recent interests. In particular, GIRAM can be seamlessly integrated with existing next POI recommendation models. Experiments on three real-world datasets demonstrate that GIRAM consistently outperforms state-of-the-art methods while maintaining high efficiency in both update time and memory consumption.

Figures (16)

• Figure 1: Illustration of traditional vs. continual next POI recommendation.
• Figure 2: Performance comparison of Static, Finetune, and Retrain methods on NYC, TKY, and CA datasets.
• Figure 3: An overview of GIRAM. It can be divided into two stages: the update stage and the deployment stage.
• Figure 4: Updating time comparison using three backbones.
• Figure 5: Memory usage comparison using three backbones.