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

Background Traditional POI recommendation systems struggle with out‑of‑town scenarios because of interest drift, ignored travel intention, and geographical gaps, leading to a classic cold‑start problem.

Proposed Solution The authors introduce the TRAINOR (TRavel‑INtention‑aware Out‑of‑town Recommendation) framework, which integrates user hometown preference, travel intention, and out‑of‑town preference into a unified model.

Key Differences from Conventional Methods

Use of a graph neural network (G‑GNN) to encode and aggregate local check‑in behavior and spatial constraints.

Modeling travel intention as a mixture of generic latent intentions (via a Neural Topic Model) and personalized intentions (via attention over local preferences).

Employing a multi‑layer perceptron (MLP) to map hometown preference to out‑of‑town preference, combined with matrix factorization and GeoConv for geographical modeling.

Model Components

Hometown Preference Modeling : Construct a directed check‑in graph for each user, embed POIs, and apply G‑GNN with attention to obtain a hometown preference embedding.

Travel Intention Discovery :

Generic intention: Unsupervised NTM learns K latent travel intentions and the POI‑intention distribution Φ.

Personalized intention: Attention over hometown preference refines the generic intention into a user‑specific intention embedding.

Out‑of‑Town Preference Modeling : Use matrix factorization to estimate out‑of‑town POI embeddings; GeoConv incorporates geographical constraints.

Preference Transfer : An MLP captures the non‑linear mapping from hometown to out‑of‑town preference.

Joint Training & Recommendation : Minimize a composite loss consisting of intention inference loss, preference loss, and transfer loss; the trained recommender scores user‑POI pairs via inner product and selects the top‑k POIs.

Experimental Validation

Datasets: Real cross‑city check‑in records (Beijing‑Shanghai, Shanghai‑Hangzhou, Guangzhou‑Foshan) collected from July to December 2019.

Baselines: TOP, SR‑GNN, LA‑LDA, EMCDR, BPR‑MF, etc., plus ablations removing intention or GeoConv modules.

Metrics: Recall@k and Precision@k, with emphasis on recall for out‑of‑town recommendation.

Results: TRAINOR consistently outperforms baselines in recall, demonstrating effective handling of cold‑start and intention modeling; qualitative analysis shows interpretable intention distributions for different user types (tourism, shopping, business).

Q&A Highlights

Check‑in graph nodes per user are ~10 on average; six‑month history is used.

G‑GNN is trained end‑to‑end as part of the composite loss; hometown preference loss is omitted because it is absorbed by the travel‑intention loss.

Transfer loss bridges the gap between local and out‑of‑town preferences, improving recommendation quality.

References

Xin et al., "Out‑of‑Town Recommendation with Travel Intention Modeling", 2021.

AAAI‑2021 paper analysis.

Kingma & Welling, "Auto‑Encoding Variational Bayes".

苏剑林, "变分自编码器（一）：原来是这么一回事".

Li et al., "Gated Graph Sequence Neural Networks".