Chicago taxi-demand model reports higher spatial equity with modest accuracy cost

A machine-learning framework evaluated on a large-scale Chicago taxi-demand dataset reduced disparities in prediction errors across demographically similar regions while retaining high overall forecasting accuracy. The study, published in Scientific Reports, reported a root mean square error (RMSE) of 2.8 and a mean absolute error (MAE) of 1.4. Its fairness mechanism produced an improvement in spatial equity of up to 50%, with a Price of Fairness (PoF) of approximately 7%.

The proposed framework has three principal components: reconstruction of missing demand data, temporal demand forecasting and fairness-aware regularization. For the first stage, the researchers combined Non-negative Matrix Factorization (NMF) with White Shark Optimization (WSO). NMF was used to identify latent spatial and temporal structures in the demand data, and the combined procedure reconstructed missing observations.

This preprocessing step supplied the forecasting stage with a completed dataset before temporal patterns were modeled. For prediction, the study used a Long Short-Term Memory (LSTM) neural network to capture temporal dependencies in taxi demand. The authors then introduced regularization based on Relative Hardship Proximity (RHP), a measure that identifies disparities in prediction errors among demographically similar regions.

The RHP mechanism penalizes unequal error distributions during model training, explicitly encouraging the network to produce more balanced results across urban areas. In the reported experiments, the hybrid system achieved the joint objective of forecasting accuracy and spatial equity. The equity gain reached as much as 50%, while the approximately 7% PoF quantified the associated trade-off in predictive performance.

The authors describe these results as evidence that structured fairness constraints can be incorporated into spatiotemporal taxi-demand models without abandoning their forecasting capabilities. The approach is notable because it treats neighborhood-level error distribution as an explicit model component rather than relying only on citywide accuracy measures. The findings remain limited to the forecasting task, Chicago dataset and experimental evaluation described in the study.

The reported metrics do not establish whether the same gains would occur in other cities, with different demand patterns or during live taxi dispatch. Nor does the experiment demonstrate changes in transportation service delivery or planning decisions. Instead, it provides a computational framework for testing whether reconstruction, temporal learning and region-sensitive error regularization can be combined in a single demand-prediction system.

The central result is a measured balance: substantially higher reported spatial equity accompanied by a comparatively small quantified cost to overall accuracy.

Subject of Research: Fairness-aware spatiotemporal taxi-demand forecasting evaluated on a Chicago dataset

Article Title: Spatial fairness in taxi demand prediction: a hybrid approach leveraging district profiles

Article References: Eslami, G., & Ghaderi, F. (2026). Spatial fairness in taxi demand prediction: a hybrid approach leveraging district profiles. Scientific Reports. https://doi.org/10.1038/s41598-026-75993-0

Image Credits: AI Generated

DOI: 10.1038/s41598-026-75993-0

Keywords: taxi demand, spatial fairness, machine learning, LSTM, missing-data reconstruction, Chicago, urban transportation, predictive accuracy, spatiotemporal modeling, fairness regularization, Spatial, fairness