Source-linked AI summary
Interdependence and Predictability of Human Mobility and Social Interactions
Manlio De Domenico, Antonio Lima, Mirco Musolesi
TL;DR
The paper addresses the limitation of predicting mobility from only an individual's past history. It applies multivariate nonlinear time-series prediction with mobility traces selected through mutual information, finding improved forecasting when correlated people or entities are included and evaluating the approach on Nokia and Cabspotting data.
Problem
Existing mobility predictors primarily use the target person's past history, leaving the predictive value of movements from socially or otherwise correlated entities to be examined.
Method
The approach adds the movement history of an entity with strongly correlated mobility patterns, selected using mutual information, to a multivariate nonlinear predictor.
Results
Mobility correlation improves movement forecasting, while correlated movement is associated with physical or distant social interactions across the Nokia and Cabspotting analyses.
Takeaways & Limitations
Mutual information can select useful additional mobility traces for prediction even when direct information about social ties is unavailable.
Takeaways & Limitations
The proposed approach has scalability constraints because multidimensional mutual-information calculations become computationally expensive as the number of users and tuple cardinality increase.
Abstract
from arXiv · showhide
Previous studies have shown that human movement is predictable to a certain extent at different geographic scales. Existing prediction techniques exploit only the past history of the person taken into consideration as input of the predictors. In this paper, we show that by means of multivariate nonlinear time series prediction techniques it is possible to increase the forecasting accuracy by considering movements of friends, people, or more in general entities, with correlated mobility patterns (i.e., characterised by high mutual information) as inputs. Finally, we evaluate the proposed techniques on the Nokia Mobile Data Challenge and Cabspotting datasets.
1. Introduction
The paper investigates whether correlated mobility patterns and social ties can improve human movement prediction beyond an individual's own history. It introduces multivariate nonlinear prediction and mutual-information-based correlation analysis, evaluated on Nokia and taxi mobility datasets.
- Human movement is predictable across geographic scales, with applications including location-aware information dissemination and targeted mobile advertising.
- The paper uses multivariate nonlinear predictors to forecast future geographic locations in longitude and latitude, contrasting them with traditional linear techniques such as ARMA.
- Mutual information quantifies correlation between mobility traces and supports selecting movement data from friends, acquaintances, or other correlated entities.
- The study reports that mobility correlation can improve movement forecasting and that correlated movement is linked to physical or distant social interactions.
- Evaluation uses GPS traces from 25 Nokia users and mobility traces from 178 taxis selected from the Cabspotting dataset.
2. Multivariate Nonlinear Time Series Prediction
The paper applies multivariate nonlinear time-series prediction to GPS mobility data, reconstructing movement dynamics from multivariate histories and testing whether nonlinear models outperform linear forecasts. The approach uses delay embedding, mutual-information-based time delays, false-nearest-neighbour analysis, surrogate testing, and local phase-space prediction.
- 2.1. Overview: The method extends nonlinear prediction from a user’s past movement history to multivariate mobility measurements, including other users when available.The state representation includes time and geographic coordinates, with altitude additionally available in the Nokia dataset.
- 2.1. Overview: Delay embedding reconstructs a phase space from time-delayed measurements, and the multivariate formulation combines delayed components from multiple time series.The embedding depends on the embedding dimensions and time delays selected for each component.
- 2.1. Overview: The nonlinear predictor searches for past states resembling the current state and averages their subsequent states to forecast future movement.This local phase-space approach assumes that similar states are likely to be followed by similar next states.
- 2.1. Overview: Linear autoregressive forecasting performs poorly for user 179, with position errors of about 3° and altitude errors generally exceeding 600 m.The linear model uses an autoregressive order of p = 24 and forecasts 1000 GPS measurements.
- 2.2.3. Analysis of Multivariate Surrogates: Mobility traces show evidence of nonlinear dynamics: false-nearest-neighbour analysis gives an optimal embedding dimension close to 30 for most Cabspotting traces, while surrogate tests reject linear stochastic dynamics.The significance exceeds 1.96 for all traces except a few exceptions representing less than 5% of the dataset.
- 2.1. Overview: For user 179, multivariate nonlinear forecasts closely agree with observed latitude and longitude, unlike the linear ARMA forecast.Figure 3 compares the true movement with ARMA and multivariate nonlinear predictions over the first 600 predictions, corresponding to about 60 hours.
3. Mutual Information and Movement Correlation
The paper uses mutual information to quantify dependence between users’ mobility traces and identify how much one user’s movement informs another’s. In this analysis, mobility is represented through geographic-position distributions, with social ties examined alongside movement correlation.
- Mutual information estimates the correlation between two multivariate mobility variables through their joint and marginal probability distributions.It equals zero when the variables are totally uncorrelated.
- For mobility analysis, each variable represents a user’s motion, with samples corresponding to geographic coordinates and PDFs measuring time spent at positions.
- The method uses mutual information to measure how much one user’s movement provides information about another user’s movement.
- In the NMDC dataset, friends or acquaintances are defined as users for whom one person appears in the other’s address book.Figure 6 compares location PDFs for socially connected and unconnected users.
4. Exploiting Movement Correlation and Social Ties to Improve Prediction Accuracy
The paper improves mobility forecasting by adding the history of another entity with strongly correlated movement to a multivariate nonlinear predictor. Across NMDC and Cabspotting, high-correlation traces improve prediction, while the NMDC analysis also connects performance gains with social ties.
- 4. Exploiting Movement Correlation and Social Ties to Improve Prediction Accuracy: The predictor combines person A’s location history with person B’s history when their mobility patterns have high mutual information.The two time-varying state vectors are juxtaposed as inputs to the multivariate nonlinear predictor.
- 4. Exploiting Movement Correlation and Social Ties to Improve Prediction Accuracy: In both datasets, adding highly correlated mobility traces consistently improves prediction accuracy.
- 4.2. Evaluation: At least one order of magnitude improvement is achieved in NMDC compared with prediction from single traces, often reaching two orders of magnitude.
- 4.2. Evaluation: More than one order of magnitude improvement is observed when the additional user is socially linked rather than socially unlinked.Social links are defined through presence in the user’s address book.
- 4.2. Evaluation: The NMDC contact–mutual-information plots show dataset-specific correlations, but the authors reject generalization because of insufficient statistics.
- 4.2. Evaluation: For Cabspotting, prediction quality is assessed over time using the error ratio e(t)/σdata, with values at or below one treated as satisfactory.The evaluation compares single-taxi, low-mutual-information, and high-mutual-information combinations.
- 4.2. Evaluation: High-correlation taxi combinations predict movements within statistical uncertainty in 80% of cases over the next 30 minutes.
5. Discussion
The discussion highlights practical uses and boundaries of mobility prediction based on correlated traces. Scalability, embedding dimensionality, and the choice of social-tie signals constrain deployment and interpretation.
- Mobility forecasting could support location-based applications by predicting where users will be and whom they may meet.The paper connects these predictions to personalised services based on users’ contexts.
- Mutual-information computation and multivariate embedding reconstruction create scalability constraints for larger user sets and higher-dimensional phase spaces.The stated complexity is O(N^n) for multidimensional mutual information, while embedding reconstruction is not feasible beyond a 40-dimensional phase space.
- No more than three mobility traces should be considered simultaneously because noisy data can produce large embedding reconstructions.This restriction follows from the reconstruction behavior described for two-coordinate mobility signals.
- Prediction quality depends on how social ties are identified, while high mobility correlation can occur even without an actual social tie.The Nokia MDC dataset lacked direct information about real or virtual social ties, and address-book presence was used as the strongest available definition.
- Non-uniform embedding and local polynomial fitting are proposed as refinements that could increase prediction accuracy.The proposed refinement uses different delays in the embedding.
6. Conclusions
The paper concludes that multivariate nonlinear time-series methods can improve movement forecasting by incorporating correlated mobility patterns. Mutual information can also select useful user pairs when social-tie information is unavailable, while mobility correlation can indicate social interactions.
- Multivariate nonlinear time-series techniques improve movement prediction by incorporating people with correlated mobility patterns.The conclusion reports this result across the paper’s analyses of user movement.
- When social-tie information is unavailable, mutual information can select user pairs to improve prediction accuracy in the Cabspotting dataset.This provides a way to use correlated traces without direct social-tie data.
- Mobility correlation measured by mutual information, together with social-tie information, can improve forecasting by exploiting other individuals’ mobility data.The conclusion frames correlation and social ties as inputs for movement forecasting.
- Mobility correlation can indicate potential physical or distant social interactions, and social interactions can likewise be associated with correlated movement.The paper states this relationship in both directions.