TL;DR
This paper investigates the potential biases introduced when using Call Detail Records (CDRs), specifically voice call data, to study human mobility. The authors propose and evaluate an "imposed sampling process" that mimics real-world data collection but with controlled parameters, finding it can offer a more representative view of mobility patterns than relying solely on voice calls.
What Problem Does This Paper Address?
Studying human mobility patterns is crucial for various applications, from urban planning to understanding disease spread. Call Detail Records (CDRs) from cellular providers offer a vast source of location data for such studies. However, CDRs, particularly those based on voice calls and SMS, are generated by user-initiated events. This means the data is inherently sparse and dependent on an individual's calling frequency, potentially leading to a biased representation of their true movement patterns. For instance, high-frequency callers might be overrepresented, and less frequent but significant locations might be missed. This paper aims to quantify this bias and explore alternative sampling methods that could provide a more accurate picture of human mobility.
Methods
The core of this paper's methodology involves comparing different data sampling strategies for human mobility. They analyze:
- Voice-call and SMS CDRs: Standard data sources used in prior research.
- Overall activity data: Combining voice, SMS, and data session records to get a more comprehensive view.
- Imposed Poisson Sampling: A novel approach where they simulate a sampling process with a controlled rate (intensity function). This process is designed to mimic the rate of voice calls but with the statistical properties of a Poisson process, which is known for its even spread. They also explore varying the intensity of this imposed sampling process.
To quantify the differences and biases, they use established mobility metrics like:
- Shannon Entropy (SU): Measures the diversity of a user's spatial footprint (how spread out their activity is across locations).
- Radius of Gyration (RG): Quantifies the range of a user's trajectory.
- Jaccard Similarity: Used to compare the overlap between sets of "significant locations" (locations accounting for over 90% of activity) derived from different data sources.
- Jensen-Shannon Divergence: A measure to quantify the difference between probability distributions, used here to compare the marginal distributions of sampled locations versus overall activity.
Datasets
The primary dataset used in this study consists of anonymized cellular voice-call, SMS, and data-session (2G and 3G) records. This data was collected from an operational CDMA 1xRTT-EVDO cellular network covering a 7,000 square mile territory in the San Francisco Bay Area. The dataset includes over a million mobile users' activity over a one-month period (July 2011). The authors do not introduce a new benchmark dataset but rather leverage this large-scale, real-world CDR data.
Key Insights and Ideas
- Voice calls are a biased sample: The study strongly suggests that relying solely on voice call CDRs can lead to a skewed understanding of human mobility. While they capture significant locations like home and work, they often miss the full spatio-temporal picture.
- Data activity is more comprehensive: Combining voice, SMS, and data session records provides a significantly more exhaustive sample of a user's spatial presence compared to voice or SMS alone.
- High-frequency callers aren't always representative: Selecting users based on high voice-call frequency can overestimate certain mobility metrics like Shannon entropy, potentially leading to a biased view of population-level mobility.
- An "imposed" sampling process can be better: A simulated Poisson sampling process, matched to the average voice-call rate, often yields results closer to the overall activity patterns than voice-call data alone. This suggests that a controlled, artificial sampling strategy can be more representative.
- Home and work locations are robust: Despite differences in overall location sampling, home and work locations are consistently identified across voice, SMS, and data activities for most users.
Key Results
- The number of distinct locations visited, as revealed by data activity, is significantly higher than that revealed by voice-call or SMS activities.
- For 30% of users, the overlap (Jaccard similarity) between significant locations identified by voice calls and data activity is as low as 10%.
- When comparing the Shannon entropy (SU) of individual users, voice-call data can incur an absolute error of 0.25 or more for over 50% of users.
- For high-activity users, the peak of the Shannon entropy distribution shifts significantly higher when comparing low-frequency versus high-frequency voice callers, suggesting an overestimation bias with high-frequency users.
- An imposed Poisson sampling process, with the same intensity as the average voice-call rate, generally performs better than voice-call sampling alone, especially for users with higher mobility and activity.
- Even with significantly reduced sampling rates (down to 1/32 of the average activity rate), an imposed Poisson process can still maintain a Jensen-Shannon divergence below 0.1 for 90% of users, indicating its robustness.
Limitations and Future Work
- Limitations: The study acknowledges that while voice calls and SMS are good at identifying home and work, they may not capture the full nuances of mobility. The paper also notes that the "significant locations" metric, while useful, doesn't account for the varying importance of these locations. The authors also mention that the Jensen-Shannon divergence is a relative measure, necessitating comparison with other processes.
- Future Work: The authors suggest that their proposed imposed sampling methodology could be directly applicable to location-sensing applications on mobile devices, aiming to sample user locations efficiently without incurring substantial overhead. They also hint at further exploring the implications of these findings for population-level mobility studies.
Summary updated