Mind the Gaps: Auditing and Reducing Group Inequity in Large-Scale Mobility Prediction
Uncovering and Rectifying Bias in Mobility Prediction Models
Our analysis of large-scale mobility data reveals significant disparities in predictive accuracy across racial and ethnic groups, highlighting the critical need for fairness-aware AI.
Executive Summary: Why Fairness in AI Mobility Matters
Neglecting fairness in AI-driven mobility predictions can lead to biased resource allocation and unequal service delivery. Our research offers a clear path to equitable outcomes.
0
TDPV Reduction (Early Stages)
0
Accuracy Difference Between Groups (Audit)
0
Accuracy Trade-off (Long-term)
Deep Analysis & Enterprise Applications
Select a topic to dive deeper, then explore the specific findings from the research, rebuilt as interactive, enterprise-focused modules.
Our audit revealed systematic disparities in predictive accuracy across racial and ethnic groups in large-scale mobility prediction models. White users consistently experienced higher accuracy, while Black and Asian users received less accurate predictions. These inequities persist across different geographical resolutions (ZCTA and county levels), indicating structural biases inherent in the data and models.
To address the lack of individual demographic data, we developed Size-Aware K-Means (SAKM), a novel clustering algorithm that partitions users in latent mobility space while enforcing census-derived group proportions. This method generated demographically grounded proxy labels, enabling group-level performance metric estimation. Building on this, we proposed Fairness-Guided Incremental Sampling (FGIS), a lightweight data acquisition strategy. FGIS prioritizes users from underrepresented or underperforming groups during data collection, balancing fairness and accuracy with a tunable tradeoff parameter (β).
FGIS successfully reduced total demographic parity violations (TDPV) by up to 40% in early sampling stages, with minimal impact on final accuracy. These improvements were most significant in low-data regimes, demonstrating the potential for fairness-aware strategies to deliver meaningful gains when data is limited or expensive. The method operates purely at the data level, requiring no access to user features or modifications to model architecture, making it a plug-and-play solution.
Quantified Impact of Fairness-Guided Sampling
0
TDPV Reduction (Early Stages)
0
Accuracy Trade-off (Long-term)
0
Initial Accuracy Gap (Groups)
Key Innovation: Size-Aware K-Means (SAKM)
SAKM Enables Demographic-Grounded Proxy Labels for Unlabeled DataSAKM is a novel clustering algorithm that extends k-means to enforce user-defined cluster size constraints, matching target group proportions derived from census data. This allows for demographic fairness evaluation without individual attributes.
Fairness-Guided Incremental Sampling (FGIS) Process
Initialize Training Set & Group Accuracies
→
Compute Group Weights (based on accuracy & representation)
→
Sample Mini-Batch from Underperforming/Underrepresented Groups
→
Add to Training Set
→
Retrain Model & Update Group Accuracies
→
Repeat Until Budget Exhausted
| Method | Pros | Cons |
|---|---|---|
| Fairness-Guided Incremental Sampling (FGIS) |
|
|
| Model-side Fairness Regularizers |
|
|
| Post-hoc Adjustments |
|
|
Case Study: Tarrant County Mobility Prediction
Scenario: We applied FGIS to a representative sub-region (Tarrant County, Texas) to validate its effectiveness. Using both MetaPath2Vec and Transformer encoder models, we simulated incremental data collection.
Outcome: FGIS successfully reduced group disparities by over 40% in early training stages with minimal impact on final accuracy. White users, initially the most favored, saw their accuracy converge more closely with other groups, demonstrating significant equity gains, especially in data-scarce scenarios.
Estimate Your Enterprise AI Fairness ROI
See how much your organization could save and how many hours could be reclaimed by implementing fairness-aware AI in your mobility and location-based services.
Your Path to Fairer AI Mobility Predictions
Our structured approach ensures a seamless integration of fairness-guided sampling and auditing into your existing AI pipelines.
Phase 1: Fairness Audit & Proxy Labeling
Conduct a comprehensive audit of existing mobility models using SAKM to generate demographically-grounded proxy labels and identify current performance disparities across groups.
Phase 2: FGIS Integration & Initial Training
Integrate the FGIS sampling strategy into your data acquisition pipeline, starting with incremental training to demonstrate early-stage disparity reduction.
Phase 3: Iterative Refinement & Monitoring
Continuously monitor group-level accuracies and TDPV. Adjust the fairness trade-off parameter (β) as needed to optimize for both fairness and overall predictive performance.
Phase 4: Scalable Deployment & Reporting
Scale the fairness-aware pipeline to full production, leveraging the data-centric approach for sustainable equity in large-scale mobility prediction systems.
Ready to Build More Equitable AI Systems?
Let's discuss how fairness-guided strategies can transform your mobility predictions, ensuring robust and responsible AI for all users.
Ready to Get Started?
Book Your Free Consultation.
Let's Discuss Your AI Strategy!
Lets Discuss Your Needs
Select a Date
Sun
Mon
Tue
Wed
Thu
Fri
Sat