Abstract
Point-of-Interest (POI) recommendation is a critical task in location-based social networks (LBSNs). Traditional methods often fail to capture the complex, non-linear patterns of human mobility and the significant influence of geographical distance. This paper introduces SAE-NAD, a novel recommendation model based on self-attentive autoencoders. By integrating a multi-dimensional attention mechanism to weigh historical check-ins and a neighbor-aware decoder to incorporate spatial constraints via an RBF kernel, SAE-NAD effectively models user preferences. Experimental results on three large-scale datasets—Gowalla, Foursquare, and Yelp—demonstrate that SAE-NAD significantly outperforms state-of-the-art baselines in terms of precision, recall, and MAP.
Introduction
Predicting a user's next destination in a Point-of-Interest (POI) recommendation system presents unique challenges that distinguish it from traditional recommendation tasks like movies or books. Human mobility is characterized by several key factors:
- Non-Linearity: Interactions between users and locations are complex and often non-linear, making traditional Matrix Factorization (MF) models insufficient.
- Unequal Importance of History: Not all past check-ins are equally representative of a user's core preferences. A routine stop should not carry the same weight as a visit to a preferred destination.
- Spatial Constraints: Physical locations are bound by distance, leading to a "clustering" effect where users tend to visit nearby locations.
The SAE-NAD Framework
The proposed model, SAE-NAD, utilizes a Stacked Autoencoder architecture enhanced with two innovative components designed to address these challenges.
Self-Attentive Encoder (SAE)
The SAE moves beyond simple averaging of user history. It employs a multi-dimensional attention mechanism that learns an "importance vector" for each POI. This allows the model to capture multifaceted user preferences—for instance, distinguishing between a preference for a restaurant's cuisine versus its atmosphere.
Neighbor-Aware Decoder (NAD)
The NAD serves as the spatial engine of the model. It incorporates geographical influence by calculating the relationship between visited and unvisited POIs using a Radial Basis Function (RBF) kernel. This ensures that recommendations are biased toward locations that are both semantically similar and geographically proximal to the user's history.
Experimental Evaluation
The model was evaluated using three extensive real-world datasets:
- Gowalla: ~1.7M check-ins (2009–2010).
- Foursquare: ~1.2M check-ins (2012–2013).
- Yelp: ~860k check-ins from the Yelp Dataset Challenge.
Data was pre-processed to filter out low-activity users and POIs, ensuring the model learned from robust behavioral patterns.
Results and Discussion
SAE-NAD was compared against a variety of baselines, including traditional MF (WRMF, BPRMF), classical POI models (RankGeoFM, IRENMF), and deep learning approaches (PACE, DeepAE).
As shown in the experimental results, SAE-NAD consistently outperforms all competing methods. For example, on the Gowalla dataset, it achieved a Precision@10 of approximately 0.070, a significant improvement over standard autoencoders (DeepAE at ~0.056) and specialized POI models (RankGeoFM at ~0.058). Ablation studies further confirmed that both the self-attention mechanism and the neighbor-aware decoder are essential for peak performance, with the spatial component providing the most substantial boost.
Conclusion and Future Work
SAE-NAD represents a significant advancement in location-intelligent deep learning. By treating geography as a decoder bias and utilizing multi-dimensional attention, it captures the nuances of human mobility more effectively than previous models. Future research could extend this work by incorporating temporal dynamics (e.g., time-of-day influences) and explicit social network ties to further refine recommendation accuracy.
Summary updated