Spatio-Temporal Check-in Time Prediction with Recurrent Neural Network based Survival Analysis

Guolei Yang, Ying Cai, Chandan K. Reddy

Abstract

Predicting the exact time of a user's check-in at a specific location is a critical yet challenging task in location-based social networks (LBSNs). Traditional models struggle with data scarcity and the "long-tail" distribution of check-ins. This paper presents Recurrent-Censored Regression (RCR), a novel framework that combines Gated Recurrent Units (GRUs) with survival analysis. By incorporating "potential visitors" through censored regression, the model effectively expands the training set and improves prediction accuracy. Experimental results on real-world datasets show a performance improvement of over 25% compared to state-of-the-art baselines.

Introduction

In the realm of location-based social networks (LBSNs), most research has focused on predicting the next location a user might visit. However, a more granular and difficult question remains: when will that visit occur? Accurate check-in time prediction has significant implications for targeted advertising, urban planning, and personalized recommendation systems.

The primary challenge in this domain is data scarcity. Most locations follow a "long-tail" distribution where a few popular venues receive many check-ins, while the vast majority have very few. Furthermore, standard regression models only learn from users who have already visited a location, ignoring "potential visitors" whose check-in event simply hasn't happened yet.

Figure 1: Overview of the check-in time prediction task, illustrating the transition from a user's trajectory to a temporal prediction.

Methodology: Recurrent-Censored Regression (RCR)

The proposed RCR framework addresses these challenges through a three-stage process:

1. User Selection and Data Enrichment

To mitigate data scarcity, the authors employ Matrix Factorization to identify "potential visitors." These are users who haven't visited the target location but exhibit similar preferences or social patterns to those who have. This step transforms the problem from a small-scale regression task into a larger-scale survival analysis task.

2. Feature Extraction

The model utilizes two categories of features:

  • User Mobility Features: These capture general behavioral patterns, such as overall activity levels and typical travel distances.
  • Check-in Features: These focus on the user's recent trajectory, identifying "precedent locations" that often serve as temporal indicators for a visit to the target venue.

3. The RCR Model and Survival Analysis

The core of the framework is a Recurrent Neural Network using Gated Recurrent Units (GRUs) to encode the temporal sequence of a user's past check-ins. This latent representation is then passed to a censored regression layer.

Unlike standard regression, which minimizes the error for observed events only, censored regression incorporates "censored" data points. In survival analysis, a censored event is one that has not occurred by the end of the observation period. By treating a check-in as a "survival event," the model can learn from both confirmed visitors and potential visitors who are likely to visit in the future.

Figure 5: Illustration of censored data in the context of user trajectories. Users $U_2$ and $U_4$ represent censored data points where the target event (check-in) has not yet occurred.

Experimental Evaluation

The RCR model was evaluated using extensive datasets from Foursquare (NYC and Tokyo) and Gowalla. The researchers compared RCR against several baselines, including Linear Regression, Deep Neural Networks (DNN), and traditional survival models like the Cox Proportional Hazards model and Accelerated Failure Time (AFT) models.

Key Results

  • Accuracy Gains: RCR consistently outperformed all baselines. The inclusion of potential visitors through censored regression led to a performance boost of approximately 26.34% to 27.17% in Time-Dependent AUC.
  • Handling Sparsity: The model proved particularly robust for "Low Popularity" locations. While standard models failed to provide meaningful predictions due to lack of data, RCR maintained high F-1 scores by leveraging the enriched dataset of potential visitors.
  • Temporal Dynamics: The use of GRUs allowed the model to capture the "momentum" of a user's day, recognizing that the sequence of locations visited is a stronger predictor of time than any single location alone.

Conclusion

The Recurrent-Censored Regression framework represents a significant step forward in spatio-temporal prediction. By reframing check-in prediction as a survival analysis problem, the authors successfully leveraged "censored" data to overcome the pervasive issue of data scarcity in LBSNs. Future work may involve integrating real-time external factors such as weather and traffic to further refine temporal accuracy.

Summary updated

Abstract

We introduce a novel check-in time prediction problem. The goal is to predict the time a user will check-in to a given location. We formulate check-in prediction as a survival analysis problem and propose a Recurrent-Censored Regression (RCR) model. We address the key challenge of check-in data scarcity, which is due to the uneven distribution of check-ins among users/locations. Our idea is to enrich the check-in data with potential visitors, i.e., users who have not visited the location before but are likely to do so. RCR uses recurrent neural network to learn latent representations from historical check-ins of both actual and potential visitors, which is then incorporated with censored regression to make predictions. Experiments show RCR outperforms state-of-the-art event time prediction techniques on real-world datasets.
BibTeX
@inproceedings{yang2018spatio,
  author = {Guolei Yang and Ying Cai and Chandan K. Reddy},
  title = {{Spatio-Temporal Check-in Time Prediction with Recurrent Neural Network based Survival Analysis}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2018},
  doi = {10.24963/ijcai.2018/413},
}
Figure