Gravity of Location-Based Service: Analyzing the Effects for Mobility Pattern and Location Prediction

Keiichi Ochiai, Yusuke Fukazawa, Wataru Yamada, Hiroyuki Manabe, Yutaka MatsuoICWSM 2020 · OpenReview

TL;DR

This paper investigates how using location-based services (LBSs) like Pokémon Go and Instagram influences user mobility patterns and improves the accuracy of predicting their future locations, especially for new places. They found that LBS users exhibit distinct behaviors, leading to up to 43.9% higher prediction accuracy for novel locations compared to non-users.

What Problem Does This Paper Address?

Predicting where a user will go next is crucial for many location-based services (LBSs). While predicting visits to frequently visited places (like home or work) is relatively straightforward, predicting visits to new or unfamiliar locations is significantly harder. This is a key challenge for applications that aim to provide timely and relevant information or recommendations. The authors hypothesize that the use of specific LBSs, which often have content tied to real-world locations (e.g., Pokémon Go's PokéStops, Instagram's scenic spots), can influence user behavior and make their movements more predictable, even for novel destinations. They aim to quantify this influence and leverage it for better location prediction.

Methods

The core of the paper involves two main studies:

  1. Mobility Pattern Analysis: The researchers conducted a quantitative analysis of check-in data to compare the mobility patterns of users of specific LBSs (Pokémon Go, Instagram) against "non-users." They analyzed metrics like the number of check-ins, the types of locations visited, the temporal distribution of visits, and the distances between successive check-ins.
  2. Location Prediction Model: Based on the insights from the mobility analysis, they proposed a novel location prediction method. This method is a variation of Collaborative Filtering (CF), specifically Lift-weighted User-based CF. The key innovation is incorporating a "lift value" that quantifies how much more likely a user is to visit a particular location because they use a specific LBS. This lift value is used to weight the contributions of similar users in the CF model, giving more importance to locations that are characteristic of a particular LBS. They also compared this to standard user-based CF and a geo-factor-enhanced CF.

Datasets

The study utilized check-in data collected from Foursquare and Instagram via Twitter. This data was collected for approximately half a year (January 4, 2018, to July 15, 2018).

  • Pokémon Go Users: Identified through keyword matching in tweets.
  • Instagram Users: Identified by tweets originating from the Instagram application.
  • "Others" / Non-users: Users who did not meet the criteria for the specific LBS users.

The data was geocoded using the Grid Square system (defined by the Japanese Statistics Bureau) with 250-meter-long sides to account for GPS noise. The authors filtered the data to ensure sufficient check-ins per user and popularity per location for robust analysis.

Key Insights and Ideas

  • LBSs Create Distinct Mobility Signatures: Users of specific LBSs don't just visit places; they visit different kinds of places and at different times than general users. This isn't just about playing the game/using the app, but about how the app's content influences their real-world movement choices.
  • "Lift" as a Predictor of LBS Influence: The concept of "lift" (how much more likely a user is to visit a place due to an LBS) is a powerful way to quantify the impact of LBSs on location choice. This is more nuanced than just looking at check-in frequency.
  • Content Matters: The nature of the LBS content (e.g., game-based like Pokémon Go vs. social sharing like Instagram) influences the types of locations users are drawn to. Pokémon Go users might be drawn to parks and nature spots, while Instagram users might favor scenic or historical sites.
  • Novelty is Key: The biggest gains in prediction accuracy are seen for novel locations, which is precisely where prediction is most challenging. This suggests LBSs can help bridge the gap in predicting unfamiliar destinations.

Key Results

  • Mobility Differences:
    • Pokémon Go users had significantly more check-ins (4.43x) than others.
  • Instagram users tended to visit large event sites, scenic, and historical places more than others.
  • Pokémon Go users tended to check into parks and nature spots/historical places more.
  • Instagram users showed distinct temporal patterns, particularly in late evening/early night.
  • Both Pokémon Go and Instagram users tended to visit more nearby locations than non-users.
  • Prediction Accuracy Improvement:
    • The proposed Lift-weighted User-based CF method significantly improved novel location predictability.
  • For Instagram users, prediction accuracy (in terms of recall@5) was up to 43.9% higher than for non-users.
  • For Pokémon Go users, prediction accuracy (in terms of recall@5) was up to 20.0% higher than for non-users.
  • The lift-weighted method generally outperformed standard user-based CF and a geo-factor-enhanced CF, especially for Instagram users.

Limitations and Future Work

  • Data Source Limitations: The check-in data is collected via Twitter, meaning it might not capture all visited locations, and users might not always accurately check-in. However, the authors argue this data is actively provided and reflects user privacy.
  • Popularity Trends: The influence of LBSs might vary with their popularity over time, which wasn't fully captured.
  • Causality vs. Association: The study shows an association between LBS usage and location bias, but not definitive causality. Future work could explore this causal link.
  • Dataset Scale and Region: While the dataset is substantial, the number of Pokémon Go users is limited, and the data is primarily from Japan. Expanding the dataset and geographical scope is suggested.
  • Future Work:
    • Develop a unified framework that incorporates all factors investigated (mobility patterns, LBS information) for location prediction.
  • Evaluate the movement effects of users of LBSs beyond Pokémon Go and Instagram.
  • Explore time-varying lift values for LBSs with rapidly changing content.
  • Investigate the implications for LBS applications like intelligent assistants, location-based advertising, and public health.
  • Potentially use this understanding to help prevent smartphone overuse by identifying patterns of LBS engagement.

Summary updated

Abstract

Predicting user location is one of the most important topics in data mining. Although human mobility is reasonably predictable for frequently visited places, novel location prediction is much more difficult. However, location-based services (LBSs) can influence users' choice of destination and can be exploited to more accurately predict user location even for new locations. In this study, we assessed the behavior difference for specific LBS users and non-users by using large-scale check-in data. We found a remarkable difference between specific LBS users and non-users (e.g., check-in locations) that had previously not been revealed. Then, we proposed a location prediction method exploiting the characteristics of check-in locations and analyzed how specific LBS usage influences location predictability. We assumed that users who use the same LBS tend to visit similar locations. The results showed that the novel location predictability of specific LBS users is up to 43.9% higher than that of non-users.
BibTeX
@inproceedings{ochiai2020gravity,
  author = {Keiichi Ochiai and Yusuke Fukazawa and Wataru Yamada and Hiroyuki Manabe and Yutaka Matsuo},
  title = {{Gravity of Location-Based Service: Analyzing the Effects for Mobility Pattern and Location Prediction}},
  booktitle = {ICWSM},
  year = {2020},
  url = {https://openreview.net/forum?id=i3YWb3K9C8},
}
Figure