CaST-POI: Candidate-Conditioned Model Improves POI Recommendation
Researchers propose CaST-POI, a candidate-conditioned spatiotemporal model for next Point-of-Interest (POI) recommendation. Unlike existing methods that compute a single user representation to score all candidates uniformly, CaST-POI uses candidates as queries to dynamically attend to user history, recognizing that the relevance of historical visits depends on the candidate being evaluated. The model also introduces candidate-relative temporal and spatial biases to capture fine-grained mobility patterns. The paper is published on arXiv under ID 2604.20845.
Key facts
- CaST-POI is a candidate-conditioned spatiotemporal model for next POI recommendation.
- Existing methods compute a single user representation from historical trajectories and score all candidates uniformly.
- CaST-POI uses candidates as queries to dynamically attend to user history.
- The model introduces candidate-relative temporal and spatial biases.
- The paper is available on arXiv with ID 2604.20845.
- The approach addresses the candidate-agnostic paradigm in current POI recommendation.
- The key insight is that the same user history should be interpreted differently for different candidates.
- The model aims to improve prediction of users' future mobility patterns.
Entities
Institutions
- arXiv