Answers are grounded in this paper and cite their sources.
Candidate-constrained point-of-interest (POI) recommendation asks a large language model (LLM) to infer a user’s next destination from an ordered trajectory and a fixed set of candidate descriptions. The model expresses its decision by generating the corresponding candidate label. Unlike item recommendation, this task must reconcile semantic preference with physical reachability and changes in visit intent. However, existing retrieval-augmented methods often select evidence by text similarity and serialize graph facts as independent sentences. They can consequently retrieve a topically relevant but geographically implausible venue and lose the relations that explain how a candidate fits the user’s current context. We propose Adaptive Spatio-Temporal Knowledge Grounding (ASKG) to carry visit-relevant POI evidence from retrieval to label generation. ASKG adjusts graph-semantic and geographic recall to the user’s activity range, gives greater weight to evidence associated with recent visits, and preserves typed relations within a fixed graph-token budget. The candidate prompt then queries these graph tokens directly. A candidate-aligned objective trains the LLM to distinguish among the valid labels. It is a weighted combination of full-vocabulary language modeling and a cross-entropy loss restricted to the n valid candidate labels. Under a candidate-constrained protocol with 10 options on Yelp and 20 on Foursquare, the reported experiments with three LLM backbones per dataset show that ASKG attains the highest MRR in all six dataset–backbone settings and the highest ACC in five. On average across the six dataset–backbone settings, ASKG achieves a relative improvement of 5.9% in MRR and 8.3% in ACC over the strongest baseline.
No takes yet. Share an insight, caveat, or question.
Tong et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: