STS: Spatial-Temporal-Semantic Personalized Location Recommendation

Wenchao Li, Xin Liu, Chenggang Yan, Guiguang Ding, Yaoqi Sun, Jiyong Zhang

TL;DR

This paper introduces STS, a novel framework for personalized location recommendations that goes beyond traditional methods by integrating spatial, temporal, and semantic information. By modeling individual user behavior with Gaussian processes and incorporating location categories, STS significantly improves recommendation accuracy, addressing the limitations of generic models.

What Problem Does This Paper Address?

Location-based social networks (LBSNs) are rich sources of data on user mobility. A key application is recommending new locations to users, enhancing their experience and enabling targeted services. While existing recommendation systems consider spatial and temporal factors, they often rely on universal models that fail to capture the diverse and heterogeneous check-in behaviors of individual users. This heterogeneity means that a one-size-fits-all approach to modeling mobility patterns is insufficient for truly personalized recommendations. The challenge is to develop a system that can adapt to each user's unique way of moving through the world and their preferences at different times.

Methods

The core of the STS framework is a personalized Gaussian Process (GP) regression model for each user. This model learns a latent function that maps temporal information to predicted future displacements. The temporal information is multifaceted, including:

  • Hour of the day: Captures daily routines.
  • Day of the week: Accounts for weekly patterns.
  • Time difference between check-ins: Reflects the pace of movement.
  • Sequence of time stamps: Models the order of activities, crucial for understanding transitions (e.g., shopping followed by a coffee break).

A modified squared exponential covariance function is used to model the non-linear relationships between these temporal features and displacements. The model parameters (covariance hyperparameters and temporal weights) are learned using stochastic gradient descent (SGD) by minimizing the negative log marginal likelihood.

To further enhance recommendations, STS incorporates a semantic-aware enhancement. This involves estimating a user's probability distribution of checking into locations of specific categories, given temporal context. This is achieved by counting category occurrences in a user's historical check-ins.

Finally, a unified recommendation framework fuses the predicted displacement (spatial-temporal influence) with the category probability (semantic-temporal influence) using a product rule to derive the final check-in probability for candidate locations. The top-N locations with the highest probabilities are recommended.

Datasets

The authors used real check-in data collected from Gowalla, a popular location-based social network. Data from five U.S. cities (Austin, Chicago, Houston, Los Angeles, and San Francisco) were utilized. The dataset includes user check-ins with geographical coordinates (longitude, latitude), time stamps, and location category information. For experiments, the data was split into 70% for training and 30% for testing.

Key Insights and Ideas

  • User Heterogeneity is Paramount: The paper strongly emphasizes that individual users exhibit vastly different mobility patterns, making universal models ineffective. This is visually demonstrated through CDF plots of user displacements.
  • Multi-faceted Temporal Information is Crucial: No single temporal feature (hour, day, time difference, sequence) is sufficient on its own. Combining these diverse temporal signals significantly improves displacement prediction accuracy.
  • Gaussian Processes for Non-linear Temporal Modeling: GPs offer a powerful way to model complex, non-linear relationships between temporal features and user displacements, outperforming linear approaches.
  • Semantic Context Matters: Incorporating location categories, especially at a finer granularity (level 2), significantly boosts recommendation performance, indicating that users have category-specific temporal preferences.
  • Synergy of Spatial, Temporal, and Semantic: The STS framework's strength lies in its unified approach, demonstrating that combining these three dimensions leads to superior results compared to models focusing on only one or two.

Key Results

  • The proposed STS framework significantly outperforms state-of-the-art location recommendation models in terms of Precision@N and Mean Reciprocal Rank (MRR) across all tested datasets.
  • For example, STS improves GeoCF and iGSLR by 15.24% and 12.41% respectively in terms of precision when only geographical influence is considered.
  • When comparing STS to other models like BaseMF, GeoCF, PTMF, iGSLR, TempMF, POI2Vec, and Distance2Pre, STS achieved an average improvement of up to 81.90% in precision and 61.93% in MRR across datasets and recommendation list sizes.
  • The performance of STS increases with more training data, but it still provides reasonable quality even with sparse data (25% training).
  • The semantic-aware enhancement, particularly using category level 2, significantly boosts performance, with STS-C2 outperforming STS-C1.

Limitations and Future Work

  • Social Influence Ignored: The paper acknowledges that social networking is an important feature but deliberately excludes it to reduce complexity. The authors note that its effect on check-in behavior is often limited in previous studies.
  • Data Sparsity: While STS handles sparsity reasonably well, it remains an inherent challenge in LBSNs that influences recommendation performance.
  • Computational Complexity: The time complexity of STS is primarily determined by matrix inversion in the Gaussian process prediction, which is O(n³). The authors suggest that this can be improved by faster matrix multiplication algorithms or approximation techniques like variational Bayesian inference.
  • Future Work:
    • Exploring more sophisticated covariance functions for GPs to better model temporal and geographic influences.
  • Adopting learning-to-rank techniques to further enhance recommendation ranking.
  • Integrating advanced NLP models (like transformers) to leverage textual information (tips, reviews) for richer semantic modeling.
  • Developing new displacement calculation methods to mitigate issues with low positioning accuracy.

Summary updated

Abstract

The rapidly growing location-based social network (LBSN) has become a promising platform for studying users’ mobility patterns. Many online applications can be built based on such studies, among which, recommending locations is of particular interest. Previous studies have shown the importance of spatial and temporal influences on location recommendation; however, most existing approaches build a universal spatial–temporal model for all users despite the fact that users always demonstrate heterogeneous check-in behavior patterns. In order to realize truly personalized location recommendations, we propose a Gaussian process based model for each user to systematically and non-linearly combine temporal and spatial information to predict the user’s displacement from their currently checked-in location to the next one. The locations whose distances to the user’s current checked-in location are the closest to the predicted displacement are recommended. We also propose an enhancement to take into account category information of locations for semantic-aware recommendation. A unified recommendation framework called spatial–temporal–semantic (STS) is introduced to combine displacement prediction and the semantic-aware enhancement to provide final top-N recommendation. Extensive experiments over real datasets show that the proposed STS framework significantly outperforms the state-of-the-art location recommendation models in terms of precision and mean reciprocal rank (MRR).
BibTeX
@article{li2020sts,
  author = {Wenchao Li and Xin Liu and Chenggang Yan and Guiguang Ding and Yaoqi Sun and Jiyong Zhang},
  title = {{STS: Spatial-Temporal-Semantic Personalized Location Recommendation}},
  journal = {ISPRS Int. J. Geo Inf.},
  year = {2020},
  doi = {10.3390/ijgi9090538},
}
Figure