The Impact of Spatial Resolution and Representation on Human Mobility Predictability

Abstract

The study of human mobility patterns is important for both understanding human behaviour, a social phenomenon and to simulate infection transmission. Factors such as geometry representation, granularity, missing data and data noise affect the reliability, validity, and credibility of human mobility data, and any models drawn from this data.This thesis discusses the impact of spatial representations of human mobility patterns through a series of analyses using entropy and trip-length distributions as evaluation criteria, Voronoi decomposition and square grid decomposition as alternative geometry representations. I further examine a spectrum of spatial granularity, from dimensions associated with social interaction, to city, and provincial scale, and toggle analysis between raw data and post-processed data to understand the impact of noisy data and missing data influence estimation. A dataset I was involved with collecting -SHED1 -featuring multi-sensor data collection over 5 weeks among 39 participants -has been used for the experiments.An analysis of the results further strengthens the findings of Song et al., and demonstrates comparability in predictability of human mobility through geometric representation between Voronoi decomposition and square grid decompositions, suggesting a scale dependence of human mobility analysis, and demonstrating the value of using missing data analysis throughout the study.ii My sincere thanks to each and every friend and colleague of mine accompanied me during those good and bad times.iii To my parents, who offered me unconditional love and support throughout the course of this thesis.iv

A refined limit on the predictability of human mobility

“…13km 2 , 40.54km 2 along with temporal quanti zation levels of 5,10,15,30,45 and 60 minutes were investi gated. These spatial and temporal quantizations reflect a similar range of quantizations to that previously considered in [3], [4]. Note that more fined-grained spatial quantization is likely to lead to inaccuracies due to GPS error and higher temporal quantizations are unlikely to provide enough data.…”

“…Previous results based on a method for computing the upper bound by [1] have suggested high predictability at or above 90% across a range of spatiotemporal quantization levels. Noting an overestimation of the bound in the formula tion in [1], due to a lack of topological constraints, an alternate formulation is provided which is shown both theoretically and empirically to provide a lower, and therefore closer to the true, upper bound providing the opportunity to revisit the studies of [1]- [4].…”

“…Later, using the Geolife dataset [10] and [3] noted a direct relationship between spatial quantization and the upper bound on predictability [3, p388-389]. Most recently, [4] again considered varying spatiotemporal quantizations but in contrast to [3] used a dataset also including indoor locations. However, compared to all previous studies this produced significantly higher upper bounds (although as their dataset is not publicly available it is not possible to replicate these results).…”

“…Considering a geo-Iocational dataset obtained from cell tower logs, heuristically corrected to conform to a Ihr sampling rate, the authors were able to provide an insight into the upper bound on predictability, with a value of 93% reported. Using their formula a number of researches have since interrogated different datasets [2]- [4] in order either extend or address some perceived weaknesses of the approach. [2] replicated the results of [1] using 14 participants via a combination of sensors while also investigating the effect of varying temporal resolution.…”

“…Since 2010, when [1] provided evidence that the probability of correctly predicting an individual's next location had an upper bound of 93%, prediction models have become significantly more advanced, focusing on predictions at increasingly fine levels of spatial and temporal granularity. This has led to increased interest in understanding mobility at these levels resulting in [2], [3] and [4] computing the upper bound using the formulation from [1] on different datasets with different properties and spatiotemporal granularities. Knowledge of the upper bounds of mobility predictors not only provides valuable insight into human behaviour, but is also of interest to the vast array of applications that benefit from accurate prediction of an individual's future location: ubiquitous advertising [5]; service provision [6]; and intelligent agents both virtual [7], [8] and physical [9].…”