[Paper Review] Mobile phone data's potential for informing infrastructure planning in developing countries

Hadrien Salat, Zbigniew Smoreda+1 more|arXiv (Cornell University)|Jul 9, 2019

Human Mobility and Location-Based Analysis21 references4 citations

TL;DR

This study demonstrates that mobile phone data can significantly improve infrastructure planning in developing countries like Senegal by enabling accurate prediction of electricity consumption and population mapping, even with limited data. By applying curve and network clustering to anonymized mobile activity, the authors achieve near-perfect prediction accuracy ($R^2$ close to 1) using only half the data, outperforming traditional census-based methods without requiring individual-level data or external training datasets.

ABSTRACT

High quality census data are not always available in developing countries. Instead, mobile phone data are becoming a trending proxy to evaluate population density, activity and social characteristics. They offer additional advantages for infrastructure planning such as being updated in real-time, including mobility information and recording visitors' activity. We combine various data sets from Senegal to evaluate mobile phone data's potential to replace insufficient census data for infrastructure planning in developing countries. As an applied case, we test their ability at predicting domestic electricity consumption. We show that, contrary to common belief, average mobile phone activity does not correlate well with population density. However, it can provide better electricity consumption estimates than basic census data. More importantly, we use curve and network clustering techniques to enhance the accuracy of the predictions, to recover good population mapping potential and to reduce the collection of required data to substantially smaller samples.

Motivation & Objective

• To assess mobile phone data's potential as a proxy for infrastructure planning in developing countries where census data are scarce or outdated.
• To evaluate whether mobile phone activity can predict domestic electricity consumption more accurately than basic census data.
• To develop a scalable, privacy-preserving method for reconstructing population and consumption patterns from limited mobile phone data samples.
• To reduce data collection costs and improve temporal tracking between censuses using clustering techniques on mobile phone activity curves.
• To validate the approach using real-world data from Senegal, including mobile call records, nighttime lights as a proxy for electricity use, and 2013 census population density.

Proposed method

• Aggregated mobile phone data from 1,298 communication towers in Senegal (2013) were spatially mapped to 1km×1km grid cells using Voronoi tessellation.
• Nighttime light intensity from NOAA’s 30-arcsecond grid served as a proxy for domestic electricity consumption.
• Population density was derived from 2013 census data, spatially intersected with Voronoi cells to align with mobile data grids.
• Curve and network clustering techniques were applied to mobile phone activity patterns to identify representative clusters and reduce data sampling needs.
• Poisson regression models were used to predict electricity consumption based on mobile activity and population density, with model fit assessed via $R^2$.
• Clustering-based sampling strategies were tested to reconstruct full census and consumption data from smaller, strategically selected samples.

Experimental results

Research questions

• RQ1 Can mobile phone data predict electricity consumption more accurately than basic census data in a developing country context?
• RQ2 Is average mobile phone activity reliably correlated with population density in Senegal, as commonly assumed?
• RQ3 Can curve and network clustering of mobile phone activity patterns enable accurate reconstruction of population and electricity consumption from significantly reduced data samples?
• RQ4 Does the inclusion of population density improve prediction accuracy beyond mobile phone activity alone?
• RQ5 Can clustering methods based solely on mobile data, without individual-level or training data, effectively recover census-level information?

Key findings

• Average mobile phone activity showed weak correlation with population density, challenging the common assumption that higher activity directly reflects higher population density.
• Mobile phone data alone provided better predictions of electricity consumption than basic census data, with $R^2$ values significantly higher than those from population density alone.
• The curve and network clustering approach enabled accurate reconstruction of census and electricity consumption data from only half the data points, achieving $R^2$ values close to one.
• The clustering method reduced required data collection to approximately one-third of the original sample size while maintaining high predictive accuracy.
• Distance matrices based on standard deviation of point-wise curve distances outperformed those based on direct correlation, and curve-based clustering proved more reliable than network-based clustering.
• Using spatial densities (per km²) instead of raw counts yielded significantly better results, indicating that activity diversity in dense areas drives demand more than raw population numbers.