Predicting visit frequencies to new places (Short Paper)

Nina Wiedemann, Ye Hong, Martin RaubalGIScience 2023 · DOI

TL;DR

This paper tackles the challenge of predicting how often a user will visit a new place, moving beyond traditional "next location" prediction. By framing it as a supervised learning problem and employing advanced neural network models like GCNs and MHSA, they demonstrate that historical mobility patterns can indeed offer insights into future visit frequencies, even for previously unvisited locations.

What Problem Does This Paper Address?

Recommending new places to users is a cornerstone of many location-based services. However, existing approaches often rely on collective user behavior or focus on predicting the immediate next location. A significant gap exists in understanding a user's potential future engagement with newly encountered locations. This is particularly challenging because, unlike established spots, there's no prior visit history for these new places. The paper highlights that a vast majority of user-visited locations are only visited once, making it difficult to predict future engagement. This paper aims to fill this gap by proposing a novel problem formulation: predicting the frequency of future visits to a new location, solely based on a single user's past mobility behavior. This is crucial for personalized services that need to anticipate user interest beyond immediate next steps.

Methods

The authors frame "visit frequency prediction" as a supervised learning task. For each user, they simulate encountering a new location by removing it from their historical data. The model then uses the remaining historical mobility data and features of the new location to predict its visit frequency.

Key techniques include:

  • Feature Engineering: Locations are represented by features such as their projected coordinates relative to the user's home, encoded location purpose, average visit start hour, and POI features.
  • Baseline Models: Simple median and k-Nearest Neighbors (kNN) approaches are used as baselines.
  • Neural Network Models:
    • MLP (Multi-Layer Perceptron): A standard fully-connected neural network.
  • MHSA (Multi-Head Self-Attention): Leverages self-attention mechanisms, similar to those used in natural language processing, to capture complex relationships in sequential data.
  • GCN (Graph Convolutional Network): Models the user's historical mobility as a graph, allowing it to learn from the structure and relationships between visited locations. The graph is embedded and then concatenated with the new location's features.

A key aspect is the use of "balanced" training data for the neural network models, where batches are sampled to ensure an equal representation of visit frequencies from 1 to 10. This addresses the inherent imbalance in real-world mobility data.

Datasets

The study utilizes data from three tracking studies:

  • Green Class 1 (GC1)
  • Green Class 2 (GC2)
  • yumuv

These datasets were collected in collaboration with the Swiss Federal Railways (SBB) and involved participants using a GNSS-tracking app and manually labeling their activities. The data was preprocessed to infer stationary staypoints and movement segments, and then used to derive visited locations via DBSCAN clustering.

For evaluation, the data was split into training and testing sets at a dataset level. The training set (Dtrain) comprised data from GC1 and yumuv, while the test set (Dtest) was sampled from GC2. To focus on rarely visited locations, only locations visited up to ten times were used as test locations.

Key Insights and Ideas

  • Predicting the "New" is Hard but Possible: The core insight is that even with limited information about a new location, a user's historical mobility patterns contain signals that can predict future visit frequency. This moves beyond simply recommending popular spots.
  • Data Imbalance is a Major Hurdle: The heavily skewed distribution of visit frequencies (many locations visited once, few visited many times) significantly complicates prediction. Addressing this imbalance, particularly through balanced sampling during training, is crucial for model performance.
  • Graph-Based Models Excel: The GCN approach, by explicitly modeling the user's mobility as a graph, shows a notable advantage in capturing complex spatial relationships and patterns, leading to better predictions.
  • Beyond Next Location: This work reframes the problem from predicting the next immediate step to understanding longer-term engagement with places, which has broader implications for personalized services.

Key Results

  • Balanced MAE Improvement: While the Mean Absolute Error (MAE) showed only marginal improvements across models, the balanced MAE revealed significant gains. The simple median baseline had a balanced MAE of 3.99, improving to 3.78 for the best kNN.
  • Neural Networks Outperform: The balanced GCN and balanced MHSA models achieved the best performance with a balanced MAE of 2.43. This indicates they can effectively learn predictive patterns from historical mobility.
  • Pearson Correlation: The Pearson correlation coefficient between predicted and ground truth visit frequencies reached up to 0.3 for the GCN and MHSA models, suggesting a moderate but meaningful correlation.
  • Qualitative Observation: Violin plots (Figure 4) show a shift in predicted frequency distribution with increasing ground truth visit frequency, even though predictions are noisy. Figure 5 illustrates the spatial distribution of predicted visit frequencies, hinting at learned spatial patterns.

Limitations and Future Work

  • Task Difficulty: The authors acknowledge that predicting visit frequencies to new locations is inherently difficult due to limited information and the stochastic nature of user decisions.
  • Data Imbalance Impact: Despite improvements, the strong imbalance in ground-truth visit frequencies remains a challenge.
  • Future Directions:
    • Exploring other models or incorporating additional contextual data could further improve performance.
  • The developed visit frequency prediction model could be used to map the probability of visits to new locations, providing insights into learned spatial visitation patterns.
  • Further research is encouraged to explore this problem formulation and its potential applications.

Summary updated

BibTeX
@inproceedings{wiedemann2023predicting,
  author = {Nina Wiedemann and Ye Hong and Martin Raubal},
  title = {{Predicting visit frequencies to new places (Short Paper)}},
  booktitle = {GIScience},
  year = {2023},
  doi = {10.4230/lipics.giscience.2023.84},
}
Figure