Source-linked AI summary
Understanding Predictability and Exploration in Human Mobility
Andrea Cuttone, Sune Lehmann, Marta C. González
TL;DR
Reported mobility-prediction accuracy varies widely, prompting the paper to investigate which factors shape next-place prediction. It examines task formulation and contextual features, finding that next-place prediction is harder than time-bin prediction and that spatial and temporal resolution affect predictive power.
Problem
Reported predictive-model accuracy ranges from 93% to under 40%, motivating investigation of the factors behind the differences.
Source: Introduction · p. 2Method
The paper investigates factors affecting next-place prediction and uses logistic regression to study the predictive usefulness of contextual features.
Source: Introduction · p. 2, Importance of contextual features · p. 12Results
Next-place prediction is the most challenging task; coarser spatial precision makes prediction easier, and temporal resolution also affects predictive power.
Source: Discussion · p. 19Takeaways & Limitations
Prediction from visited-place histories fails on explorations, which the paper distinguishes from predicting returns to known places.
Source: Discussion · p. 20A full solution for predicting the next explored place is beyond the scope of this work.
Source: Discussion · p. 20
Abstract
from arXiv · showhide
Predictive models for human mobility have important applications in many fields such as traffic control, ubiquitous computing and contextual advertisement. The predictive performance of models in literature varies quite broadly, from as high as 93% to as low as under 40%. In this work we investigate which factors influence the accuracy of next-place prediction, using a high-precision location dataset of more than 400 users for periods between 3 months and one year. We show that it is easier to achieve high accuracy when predicting the time-bin location than when predicting the next place. Moreover we demonstrate how the temporal and spatial resolution of the data can have strong influence on the accuracy of prediction. Finally we uncover that the exploration of new locations is an important factor in human mobility, and we measure that on average 20-25% of transitions are to new places, and approx. 70% of locations are visited only once. We discuss how these mechanisms are important factors limiting our ability to predict human mobility.
Introduction
Reported mobility-prediction accuracy varies widely, and this paper investigates the factors behind that variation while examining exploration, contextual features, and a longitudinal location dataset.
Introduction
- Prior studies report predictive accuracy ranging from as high as 93% to under 40%, motivating investigation of the factors behind those differences.Source: Introduction · p. 2
- The paper identifies upper-limit versus actual prediction, task formulation, spatial resolution, and temporal resolution as factors shaping reported predictability.It contrasts predicting a next location with identifying the location in the next time-bin.Source: Introduction · p. 2
- Exploration and locations visited only once are identified as key factors limiting individual mobility-prediction accuracy.Source: Introduction · p. 2
- The paper measures the predictive power of contextual features, including social proximity, time, calls, and SMS.Source: Introduction · p. 2
- More than 400 users contribute to a novel, longitudinal, high-precision location dataset for studying human-mobility predictability.Source: Introduction · p. 2
Related work
Prior work estimated high predictability limits and reported varied performance from actual prediction models. The paper positions exploration and other intrinsic mobility characteristics as underexamined factors in next-place prediction.
Related work
- For a majority of users, Song et al. estimated an upper predictability bound of 93%; later work reported tighter bounds of 81–85% and 85%.The tighter bounds came from studies considering spatial reachability and post-earthquake mobility, respectively.Source: Related work · pp. 2–3
- Markov-chain results varied across settings: over 90% accuracy for daily county-level prediction, and median accuracy of about 65–72% for campus WiFi traces.These results concern different datasets and prediction goals.Source: Related work · p. 3
- Naive Bayes studies reported approximately 50% accuracy with time and location features and approximately 60% with additional SMS, call, and Bluetooth-proximity features.Source: Related work · p. 3
- Other explored approaches include nonlinear time series, principal component analysis, Gaussian mixtures, and dynamic Bayesian networks.Source: Related work · p. 3
- In 53 weeks, users visited an average of 200 unique locations, 70% visited only once; exploration prediction reached at best 41% accuracy despite most trips involving recurrent locations.The paper identifies exploration as a factor limiting next-place prediction.Source: Related work · p. 3
Materials and Methods
This section describes the smartphone location histories and modeling approach used to examine mobility prediction across task formulations and contextual features. It also covers a separate exploration-prediction task and comparisons of features relevant to that task.
Data description
- The analysis retained 454 users with at least three months of contiguous data, selecting each person’s longest period with samples in at least 90% of 15-minute bins.Location samples were collected every 15 minutes, with median accuracy of approximately 20 meters; more than 90% were reported accurate to better than 40 meters.Source: Data description · p. 4
next-cell prediction
- The next-cell predictability limit peaked around 0.95, while Markov accuracy peaked around 0.7 and closely tracked the stationary strategy (r = 0.993, p < 0.001).The stationary strategy predicts that the user remains in the current cell, suggesting that much of Markov’s predictive power comes from self-transitions.Source: next-cell prediction · pp. 6–7, Figure 4 · p. 7, Figure 6 · p. 9
- Increasing grid-cell size improved Markov accuracy, while increasing time-bin duration reduced it as self-transitions became less frequent.The study compared 50 m, 500 m, and 5000 m cells, and 15-, 30-, and 60-minute time bins.Source: next-cell prediction · pp. 7–8, Figure 7 · p. 10, Figure 8 · p. 11
next-place prediction
- For next-place prediction, stops were detected using a 50-meter threshold and grouped into places with DBSCAN using a 50-meter distance and min pts = 2.Stops shorter than or equal to 15 minutes were filtered out, and consecutive stops at the same place were merged.Source: next-place prediction · p. 9
- Next-place accuracy was around 0.3 for toploc and 0.4 for Markov, lower than in the next-cell formulation.The authors characterize next-place prediction as the more difficult task; the next-cell toploc and Markov accuracies peaked around 0.4 and 0.7, respectively.Source: next-cell prediction · pp. 6–7, next-place prediction · p. 12, Figure 12 · p. 14
Importance of contextual features
- Location and time features outperformed the Markov baseline; call and SMS features had little predictive power, and all features performed similarly to location plus hour of week.The logistic regression models evaluated individual features and a combined feature set for next-place prediction.Source: Importance of contextual features · pp. 12–14
Understanding the set of location states
- About 70% of places were visited only once, and 20–25% of stops were explorations, making unseen places a major prediction challenge.Models drawing predictions from previously visited places cannot predict new, unseen symbols; the pool of possible places also grows over time.Source: Understanding the set of location states · pp. 14–15
Exploration prediction
- The exploration task predicts whether the next stop is new or a return, using logistic regression with added current- and prior-exploration features; the study uses F1 because the classes are imbalanced.The proposed baseline guesses exploration with probability approximately 0.2; models are fit per user online.Source: Exploration prediction · pp. 16–17
- The best-performing exploration model reached average precision of 0.3 and recall of 0.65, showing that exploration prediction remains challenging.Source: Exploration prediction · p. 18
- The model with the most complete feature set outperformed the others, while current location performed best among single-feature models.The authors suggest that places such as public transport hubs can act as gateways for exploration.Source: Exploration prediction · p. 18
- Time-related features and the “explore now” and “explore before” features also performed well, consistent with time-specific and bursty exploration.The passage notes exploration may occur at particular times of day or week and may show burstiness.Source: Exploration prediction · p. 18
Discussion
Prediction accuracy depends on how the task and data resolution are defined: predicting a next time-bin is easier than predicting the next transition. Ongoing exploration adds unseen places to the prediction problem, while demographic effects and the value of additional information channels remain open questions.
Discussion
- About 4 new places are explored per user each week, and the possible-place pool reaches almost 200 after one year.The average number of new places explored per week does not decrease over time.Source: Figure 16 · p. 19, Figure 17 · p. 20
- Predicting the next time-bin location is much easier than predicting the next transition, while spatial and temporal resolution also affect predictive power.The paper identifies problem formulation as central when interpreting predictability results.Source: Discussion · pp. 18–19
- Demographics may contribute to performance differences, but the student population and its changing class schedules limit conclusions until the link is studied further.The authors say more work is needed to conclusively link demographics and predictability.Source: Discussion · p. 19
- Exploration matters for next-place prediction because models restricted to visited places fail on new locations, and full next-exploration prediction is beyond this work's scope.The paper distinguishes predicting explorations from predicting returns to known places.Source: Discussion · pp. 19–20
- Location history may not suffice for accurate next-place prediction; the authors speculate that email, social media, calendars, or class schedules may be needed.Schedules change, plans are canceled, and people run late despite regularities in visit sequences and temporal patterns.Source: Discussion · p. 20