Geo-ALM: POI Recommendation by Fusing Geographical Information and Adversarial Learning Mechanism

Wei Liu, Zhi-Jie Wang, Bin Yao, Jian Yin

Abstract

Point-of-Interest (POI) recommendation is a critical task in location-based social networks (LBSNs). Traditional methods often suffer from the "easy negative" problem and coarse geographical modeling. We propose Geo-ALM, a framework that integrates Generative Adversarial Networks (GANs) with multi-layered geographical data. By employing a generator to identify "critical" negative samples and a discriminator that incorporates both POI-level and region-level geographical constraints, Geo-ALM significantly improves recommendation accuracy. Experimental results on Foursquare and Gowalla datasets demonstrate that Geo-ALM outperforms state-of-the-art baselines.

Introduction

In location-based social networks (LBSNs) like Yelp or Foursquare, users typically visit only a small fraction of available POIs. Learning user preferences usually involves pairwise ranking, assuming a user prefers a visited location over an unvisited one. However, two major challenges persist:

  1. The "Easy" Negative Problem: Randomly sampled unvisited POIs often serve as "easy" negatives (e.g., a coffee shop in a different city), providing little informative signal. Models require "critical" or "hard" negatives—places a user might have visited but didn't—to learn nuanced preferences.
  2. Coarse Geographical Modeling: Most models treat geography through a single lens (e.g., coordinates), failing to capture the hierarchical nature of human movement, which often involves selecting a neighborhood before a specific venue.
Training with critical negatives in geographical space.

Methodology: The Geo-ALM Framework

The proposed Geographical Information Based Adversarial Learning Model (Geo-ALM) addresses these issues through an adversarial game between two components:

Adversarial Components

  • The Generator (G): Instead of random sampling, the generator learns to select unvisited POIs that are likely to confuse the discriminator. Due to the discrete nature of POI selection, Policy Gradient reinforcement learning is used for training.
  • The Discriminator (D): This component aims to distinguish between a user's actual visits and the realistic negative samples provided by the generator.
The Geo-ALM architecture showing the interaction between the discriminator (D) and generator (G) with geographical constraints.

Two-Granularity Geographical Modeling

To enhance embeddings, Geo-ALM incorporates geographical features at two levels:

  • POI Level: A Skip-gram inspired approach ensures that physically proximal POIs share similar latent features.
  • Region Level: The map is divided into grids, modeling preference as a two-step process: choosing a region, then a POI within it. This helps mitigate data sparsity by leveraging broader regional activity.

Experimental Evaluation

The model was evaluated on two public datasets: Foursquare (Singapore) and Gowalla (California/Nevada). Both datasets are extremely sparse (density < 1.5%).

Key Results

  • Performance Gains: Geo-ALM consistently outperformed state-of-the-art baselines like BPR, Rank-GeoFM, and GeoIE. On Foursquare, it achieved a 7% relative improvement over the strongest competitor.
  • Precision and Recall: Superior performance in Precision@n and Recall@n (n=3, 5, 10) indicates better ranking of relevant POIs.
  • Ablation Study: Removing either the adversarial mechanism or geographical features led to performance degradation, validating the fusion of these techniques.

Conclusion and Future Work

Geo-ALM demonstrates that combining adversarial training with multi-level geographical information significantly improves POI recommendation. By focusing on "critical" negatives and hierarchical spatial contexts, the model achieves more precise user and POI embeddings. Future work includes incorporating temporal dynamics and social influence to further refine recommendations, as well as addressing hyperparameter sensitivity regarding region size and neighbor count.

Summary updated

Abstract

Learning user’s preference from check-in data is important for POI recommendation. Yet, a user usually has visited some POIs while most of POIs are unvisited (i.e., negative samples). To leverage these “no-behavior” POIs, a typical approach is pairwise ranking, which constructs ranking pairs for the user and POIs. Although this approach is generally effective, the negative samples in ranking pairs are obtained randomly, which may fail to leverage “critical” negative samples in the model training. On the other hand, previous studies also utilized geographical feature to improve the recommendation quality. Nevertheless, most of previous works did not exploit geographical information comprehensively, which may also affect the performance. To alleviate these issues, we propose a geographical information based adversarial learning model (Geo-ALM), which can be viewed as a fusion of geographic features and generative adversarial networks. Its core idea is to learn the discriminator and generator interactively, by exploiting two granularity of geographic features (i.e., region and POI features). Experimental results show that Geo- ALM can achieve competitive performance, compared to several state-of-the-arts.
BibTeX
@inproceedings{liu2019geo,
  author = {Wei Liu and Zhi-Jie Wang and Bin Yao and Jian Yin},
  title = {{Geo-ALM: POI Recommendation by Fusing Geographical Information and Adversarial Learning Mechanism}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2019},
  doi = {10.24963/ijcai.2019/250},
}
Figure