Research map / Learning theory and algorithms / Fairness

FairTP: A Prolonged Fairness Framework for Traffic Prediction

AAAI 2025 · Learning theory and algorithms › Fairness · approach: Fairness auditing and regularization · 3 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes FairTP, a plug-in framework that makes graph-based traffic prediction models fairer across city regions and over time via two new fairness metrics (region-based static fairness and sensor-based dynamic fairness) and a state-guided balanced sampling strategy.

Problem

Traffic prediction models are inaccurate in regions with few sensors, producing unfair regional decisions, and static fairness at individual time points is not preserved over time.

Approach

Defines region-based static fairness (RSF) as the MAPE gap between regions at each time point and sensor-based dynamic fairness (SDF) over a period T_d, using a state identification module that labels sensors as "benefit" or "sacrifice" from MAPE thresholds; a state-guided balanced sampler then oversamples sacrificed sensors and balances sampling across regions.

Key results

• On SD, FairTP-AGCRN reduces MAE from 19.215 to 14.798 and RSF from 1.337 to 1.002 relative to AGCRN.
• On HK, FairTP-GWNET lowers RSF from 1.688 to 0.835 while MAE improves from 2.189 to 2.132.
• FairTP-D2STGNN outperforms FairST and SA-Net mitigation baselines on both datasets in RSF (e.g., HK RSF 0.82 vs 1.86 for FairST-D2STGNN).

Measured effects reported in the paper

Surprising findings

• Removing the static RSF loss (noS) lowers the RSF-based fairness gains in some cases while the SDF term dominates the dynamic fairness effect, with SDF values varying by orders of magnitude across ablations (e.g., noD SDF 344.12 on SD).
• Using fixed stratified sampling instead of the state-guided sampler yields slightly better MAE for some backbones (e.g., noAS MAE 1.78 on HK for AGCRN), indicating an accuracy-fairness trade-off.

Limitations

• Sensor state labels rely on a MAPE threshold taken from a previously trained base model, which adds a dependency on that model's training run.
• Results are reported only on two datasets (HK and SD) and region partitions are fixed at 13 and 12 regions.
• Some reported accuracy degradations occur (e.g., FairTP-DCRNN MAE 2.411 vs 2.353 on HK; FairTP-ASTGCN MAE 2.110 vs 2.005).

Experimental setup

Datasets: HK Didi taxi trajectory dataset SD PeMS dataset

Benchmarks: HK Didi dataset SD PeMS dataset

Metrics: MAE RMSE MAPE RSF SDF

Baselines: DCRNN AGCRN GWNET ASTGCN DSTAGNN DGCRN D2STGNN FairST SA-Net

Builds on: DCRNN AGCRN D2STGNN

Closest papers by meaning

• On the Maximal Local Disparity of Fairness-Aware Classifiers (ICML 2024 · similarity 0.83)
• Geometry of Relaxed Fair Regression: A Unified Framework for Aware and Unaware Settings (NeurIPS 2026 · similarity 0.83)
• Intersectional Fairness via Mixed-Integer Optimization (NeurIPS 2026 · similarity 0.83)
• Adapting Static Fairness to Sequential Decision-Making: Bias Mitigation Strategies towards Equal Long-term Benefit Rate (ICML 2024 · similarity 0.83)
• FairTrade: Achieving Pareto-Optimal Trade-Offs between Balanced Accuracy and Fairness in Federated Learning (AAAI 2024 · similarity 0.83)
• TraffiDent: A Dataset for Understanding the Interplay Between Traffic Dynamics and Incidents (NeurIPS 2025 · similarity 0.82)