BiasBuster: a Neural Approach for Accurate Estimation of Population Statistics using Biased Location Data.

Area of Science:

• Data Science
• Computational Social Science
• Machine Learning

Background:

• Mobile device location data is widely used for urban mobility, business insights, and public health policy.
• These datasets often suffer from population bias, with certain communities over or underrepresented.
• Aggregate statistics derived from biased data lead to inaccurate population representations and disproportionately affect marginalized groups.

Purpose of the Study:

• To address the challenge of generating accurate population statistics from biased mobile location data.
• To evaluate the effectiveness of traditional statistical debiasing methods.
• To introduce and validate a novel neural network approach for bias correction.

Main Methods:

• Proposed BiasBuster, a neural network model leveraging correlations between location characteristics and population statistics.
• Conducted extensive experiments using real-world location data.
• Compared BiasBuster's performance against traditional statistical debiasing techniques.

Main Results:

• Statistical debiasing methods often fail to significantly improve accuracy.
• BiasBuster demonstrated substantial improvements in estimating population statistics.
• Accuracy was enhanced up to twofold generally and threefold for underrepresented populations.

Conclusions:

• Biased location data poses significant risks for policy-making and research.
• BiasBuster offers a robust solution for obtaining more accurate population statistics from mobile location data.
• The findings highlight the potential of machine learning in mitigating data bias for equitable insights.