Source-linked AI summary
STP-UDGAT: Spatial-Temporal-Preference User Dimensional Graph Attention Network for Next POI Recommendation
Nicholas Lim, Bryan Hooi, See-Kiong Ng, Xueou Wang, Yong Liang Goh, Renrong Weng, Jagannadan Varadarajan
TL;DR
Next POI recommendation is limited by methods that learn POI relationships mainly within individual users’ sequences rather than across users globally. STP-UDGAT combines personalized exploitation, global spatial-temporal-preference exploration, selective user-to-user learning, and random-walk masked attention, outperforming baselines and state-of-the-art methods across six real-world datasets.
Problem
RNN-based next-POI methods mainly learn POI-POI relationships from independent user sequences, limiting direct global cross-user learning.
Method
STP-UDGAT balances personalized preference exploitation with global spatial-temporal-preference exploration, selective user learning, and random-walk masked self-attention for higher-order neighbours.
Results
Experiments across six real-world datasets show that STP-UDGAT outperforms baseline and state-of-the-art methods.
Takeaways & Limitations
The approach jointly uses local and global views to learn POI-POI and user-user relationships for next-POI recommendation across LBSN, transportation, and terrorism applications.
Takeaways & Limitations
The GAT attention coefficients assume scalar coefficients adequately represent each neighbour’s whole vector representation.
Abstract
from arXiv · showhide
Next Point-of-Interest (POI) recommendation is a longstanding problem across the domains of Location-Based Social Networks (LBSN) and transportation. Recent Recurrent Neural Network (RNN) based approaches learn POI-POI relationships in a local view based on independent user visit sequences. This limits the model's ability to directly connect and learn across users in a global view to recommend semantically trained POIs. In this work, we propose a Spatial-Temporal-Preference User Dimensional Graph Attention Network (STP-UDGAT), a novel explore-exploit model that concurrently exploits personalized user preferences and explores new POIs in global spatial-temporal-preference (STP) neighbourhoods, while allowing users to selectively learn from other users. In addition, we propose random walks as a masked self-attention option to leverage the STP graphs' structures and find new higher-order POI neighbours during exploration. Experimental results on six real-world datasets show that our model significantly outperforms baseline and state-of-the-art methods.
1 INTRODUCTION
Next POI recommendation must model nonlinear user preferences while overcoming the local, user-sequence view of many RNN approaches. STP-UDGAT addresses this by combining personalized exploitation with global spatial-temporal-preference exploration and reports superior results across six real-world datasets.
- Next POI recommendation supports LBSN personalization and applications including ride-hailing pickup or drop-off prediction and terrorism-related POI prediction.
- RNN-based methods model sequential dependencies and incorporate spatial or temporal factors, but primarily learn POI-POI relationships within individual users’ visit sequences.
- STP-UDGAT learns POI-POI relationships from local and global views using spatial, temporal, and preference factors.
- The model balances personalized preference exploitation with exploration of global spatial-temporal-preference neighbourhoods and selectively learns from similar users.
- Random-walk masked self-attention traverses graph structure to attend to relevant higher-order POI neighbours beyond first-order neighbours.
- Experiments on six real-world datasets across LBSN, terrorism, and transportation show that STP-UDGAT outperforms baseline and state-of-the-art methods.
2 RELATED WORK
Related work progresses from collaborative filtering and sequential models toward graph representation learning for recommendation. STP-UDGAT extends this direction by using GAT with spatial-temporal-preference structure for global and local next-POI relationships.
- Earlier next-POI methods used matrix factorization, Markov chains, personalized ranking, latent spaces, and content-aware POI embeddings.
- Figure 1 depicts common shopping-mall POIs across different users’ visit sequences who share a shopping-mall preference.
- RNN-based approaches model sequential dependencies in user visit sequences, including spatial-temporal intervals and long- and short-term preferences.
- Graph-based recommendation methods use global preference and local session factors, while GAT applies masked self-attention to graph neighbourhoods.
- STP-UDGAT is presented as the first next-POI work to use GAT with spatial-temporal-preference factors, POI-POI and user-user relationships, and random-walk attention for higher-order neighbours.
- Unlike GE, STP-UDGAT learns graph representations with GAT and balances explore-exploit trade-offs across local and global views.
3 PRELIMINARIES
The task ranks candidate POIs for a user based on historical visits, while GAT computes hidden vertex representations by attending to graph neighbours and aggregating their projected features.
- The next-POI task uses a user’s historical POI visits to produce an ordered recommendation set in which the next visited POI ranks highly.
- A GAT layer computes a hidden representation for each vertex by applying masked self-attention to its graph neighbourhood.
- The layer accepts a vertex representation and an adjacency-defined neighbourhood, whose closed neighbourhood contains the vertex and its first-order neighbours.
- GAT computes attention coefficients for neighbouring representations, with values from 0 to 1 indicating how strongly each neighbour is attended.
- Softmax produces attention coefficients that form a probability distribution, and the learned coefficients weight the vertex-neighbour representations in the output aggregation.
4 APPROACH
The approach combines personalized historical preferences with globally constructed spatial, temporal, and preference POI graphs. DGAT-based exploration, random-walk higher-order neighborhoods, and user-user attention are fused to balance exploitation and exploration for next-POI prediction.
- 4.1 DGAT: DGAT assigns a separate attention coefficient to each embedding dimension instead of using one scalar coefficient for an entire neighbor representation.The dimensional coefficients are applied through a Hadamard product when aggregating neighbor representations.
- 4.2 PP-DGAT: A personalized preference graph connects every pair of a user’s historical training POIs, enabling self-attention over that user’s local preferences.The graph is complete and undirected, and its vertices are the user’s historical training POIs.
- 4.2 PP-DGAT: PP-DGAT uses the previous POI and all available historical POIs to produce a probability distribution over the full POI set for next-POI recommendation.The distribution is obtained by projecting the DGAT representation to |V| classes and applying softmax, then ranking POIs by probability.
- 4.3 STP-DGAT: Spatial and temporal graphs encode global POI relationships, including nearest-neighbor proximity and chronological transitions between visits.The spatial graph links each POI to its top five nearest POIs, while the temporal graph links chronologically successive POI visits using averaged time intervals.
- 4.3 STP-DGAT: Random-walk masked self-attention explores higher-order STP neighbors beyond first-order adjacency, while a linear layer fuses the resulting exploration representation with PP-DGAT exploitation.The fused representation balances newly explored unvisited POIs with the user’s historical preferences.
- 4.3 STP-DGAT: STP-UDGAT extends the model with user semantics so users can attend to other similar users during recommendation.The final variant incorporates user representations through concatenation before the output projection.
5 EXPERIMENTS
Experiments across six datasets and three application domains evaluate STP-UDGAT against established baselines, variants, and ablations. Results generally favor combining personalized exploitation with global STP exploration, though the strongest variant varies by dataset and metric.
- Datasets and setup: Six datasets span LBSN, terrorism, and transportation applications, with next-POI prediction evaluated from sequential user records.The datasets include four LBSN collections, the Global Terrorism Database, and a Southeast Asian ride-hailing trajectory dataset.
- Datasets and setup: The experiments use Acc@K for K∈{1, 5, 10, 20} and MAP to evaluate ranked recommendation quality.Acc@K checks whether the ground-truth POI appears in the top K, while MAP evaluates the quality of the full recommendation list.
- Overall results: Our variants significantly outperform baselines and LSTPM across all metrics, with a highest average relative improvement of 8.33% for Acc@10.Across individual datasets, one of the three proposed variants always achieves the best result for every metric.
- Dataset-specific results: STP-UDGAT is best on all metrics for Gowalla, Foursquare-Global, and Foursquare-SG, whereas PP-DGAT-Skip is best on Brightkite and GTD.Transport-SEA instead gives STP-UDGAT the best Acc@1 and STP-DGAT the best remaining metrics.
- Robustness: On Foursquare-SG, STP-UDGAT matches U-TOP on Acc@1 but leads on the remaining metrics; under cold start, it surpasses all baselines and LSTPM on Foursquare-Global Acc@1.The cold-start preprocessing retains POIs visited by more than one user and users with fewer than ten visits.
- Ablation study: Ablations show that combined STP graphs, random-walk attention with adjacency attention, dimensional attention, and explore-exploit fusion each improve performance.Exploration alone outperforms exploitation of historical POIs, while combining both yields a large further increase; performance also rises as more unvisited POIs are explored.
6 CONCLUSION
The paper presents STP-UDGAT for next POI recommendation and reports effectiveness across six real-world datasets spanning LBSN, transport, and terrorism applications.
- STP-UDGAT is presented as a novel explore-exploit model for the next POI recommendation task.
- Experiments on six real-world datasets support the approach's effectiveness across LBSN, transport, and terrorism applications.
- Future work will study how transportation pick-up points can support the recommendation task.