Travel Intention‑Aware Out‑of‑Town POI Recommendation (TRAINOR) Framework

The paper addresses the classic cold‑start problem in out‑of‑town POI recommendation, where users lack historical records for unfamiliar destinations and their travel intentions are complex.

Key challenges: interest drift between hometown and destination, ignored travel intention, and geographical gaps across cities.

Proposed solution (TRAINOR): a three‑stage framework that (1) models hometown preference using a graph‑based G‑GNN on users' local check‑in graphs, (2) discovers travel intention via an unsupervised neural topic model (NTM) to learn a distribution over K latent intentions and refines it with user‑specific attention, and (3) models out‑of‑town preference with matrix factorization and GeoConv to capture geographic constraints.

Preference transfer: a multi‑layer perceptron (MLP) maps hometown preference embeddings to out‑of‑town embeddings, jointly optimizing intention inference loss, preference loss, and transfer loss.

Training & recommendation: the combined loss is minimized end‑to‑end using Adam (lr=0.001, L2=1e‑5). At inference, the top‑k POIs with highest scores are recommended.

Experiments: Real cross‑city check‑in data (Beijing‑Shanghai, Shanghai‑Hangzhou, Guangzhou‑Foshan) collected via crawlers. Baselines include TOP, SR‑GNN, LA‑LDA, EMCDR, BPR‑MF, etc. TRAINOR consistently outperforms baselines in recall@k and shows competitive MAP, especially when the travel‑intention module is active.

Qualitative analysis: case studies reveal that inferred intentions differentiate leisure (e.g., tourism) from business trips, and POI intention distributions align with functional categories (scenic spots vs. facilities).

Q&A highlights: average check‑in graph nodes ≈10 per user, six‑month history used, G‑GNN trained jointly with other losses, GeoConv contributes modestly, and transfer loss bridges hometown‑destination preference gaps.

Conclusion: By jointly modeling hometown preference, travel intention, and out‑of‑town preference, TRAINOR effectively mitigates cold‑start issues and improves recommendation recall for cross‑city travel scenarios.