Research & Papers

CaST-POI: Candidate-Conditioned Spatiotemporal Modeling for Next POI Recommendation

New model reads your history differently for each place you might go next.

Deep Dive

Next Point-of-Interest (POI) recommendation is critical for location-based services like maps, ride-sharing, and travel apps. Traditional models compute one user representation from historical check-ins and apply it uniformly to score all possible destinations. This candidate-agnostic approach misses a key nuance: the relevance of a past visit depends on the specific place being considered. For example, a coffee shop visit matters more for predicting a café than a gym.

To address this, researchers propose CaST-POI, a model that conditions its interpretation of user history on each candidate POI. It uses a candidate-conditioned sequence reader where the candidate acts as a query to dynamically attend to relevant parts of the user's trajectory. Additionally, it introduces candidate-relative temporal and spatial biases to model fine-grained mobility patterns. Experiments on three benchmark datasets (e.g., Foursquare, Gowalla) show CaST-POI consistently beats state-of-the-art baselines, with particularly strong gains when the candidate pool is large. The code is open-source, enabling further research and deployment.

Key Points
  • CaST-POI dynamically interprets user history based on each candidate POI, not a single representation.
  • Uses candidate-relative temporal and spatial biases to capture nuanced mobility patterns.
  • Outperforms state-of-the-art methods on three benchmark datasets, especially with large candidate pools.

Why It Matters

CaST-POI makes location recommendations smarter by treating each destination as a unique query, improving accuracy for apps.