Fairness-aware Spatio-temporal Prediction for Cities

Abstract

Equitable prediction is of critical importance to urban applications such as transportation resource allocation and predictive policing. Decisions based on unfair predictions may lead to inequitable service distribution or impose disproportionate impact on underrepresented minorities. Machine learning based spatio-temporal prediction models have been widely adopted in urban settings, but few of them have built fairness into their design. This dissertation is a pioneering work to explore a suite of fairness-aware spatio-temporal prediction methods for cities, including measuring fairness for urban applications (metrics), designing fairness-aware spatio-temporal prediction algorithms (algorithms), learning bias-free data representations (data), and evaluating fairness-aware systems for real-world applications (applications). Specifically, I propose FairST, a fairness-aware spatio-temporal prediction model based on 3D convolutional neural network. A key feature of FairST is the integration of fairness regularizers to the model to encourage equitable prediction. I also propose two fairness metrics that measure equity gaps between social groups for urban mobility systems. Experiments on two real-world new mobility datasets demonstrate that FairST is able to close more than 80% of fairness gap while achieving better accuracy than state-of-the-art but fairness-oblivious baseline methods. Further experiments show that FairST is able to reduce unfairness for multiple attributes without sacrificing much accuracy. I propose an unsupervised algorithm framework to learn fair, accurate, and reusable (FAR) data representations, the EquiTensors, for heterogeneous and multi-dimensional urban datasets. Experiments with 23 input datasets and 4 real applications suggest that EquiTensors could help mitigate the effects of the sensitive information embodied in the biased data. Meanwhile, applications using EquiTensors outperform models that ignore exogenous features and are competitive with "oracle" models that use hand-selected datasets. EquiTensors can be trained and released by government agencies or trusted data brokers over both public open data and unreleased data. It presents a novel way to allow downstream applications a means of improving accuracy, avoiding data discovery and pre-processing, and limiting their exposure to new sources of discriminatory bias. This dissertation will make methodological contributions to urban data science and machine learning research. The proposed methods will inform the development of fairness assessment measures and bias-removal strategies for stakeholders such as public resource/service distributors and government agencies, allowing for intelligent and responsible decision-making that benefits all citizens.