Overview

This study investigates socioeconomic and demographic biases embedded in New York City’s ride-hailing pricing algorithms, combining trip data from the NYC Taxi and Limousine Commission (TLC) with neighborhood-level demographic insights. Our analysis reveals that economically vulnerable communities, particularly neighborhoods with high elderly poverty rates, disproportionately favor private rides over shared options—a preference that exacerbates financial strain. By comparing conventional algorithmic pricing models with fairness-aware alternatives, we demonstrate that adjustments for equity can reduce fare disparities by up to 18%, albeit with a modest trade-off in predictive accuracy (8–12%). These findings underscore the tension between profit-driven algorithmic design and equitable urban mobility, urging stakeholders to rethink how fairness is prioritized in automated systems.

Background

Transportation and real estate systems have long played a significant role in the economic livelihood of citizens, especially in New York City. In the wake of the Great Depression, Robert Moses' "Urban Renewal Plan cut swathes through vibrant Black and Jewish communities, forcing residents to relocate to areas with less work. 'Redlining', where banks denied loans in ethnic neighborhoods, stymied the economic development of many non-white residents. The effects of these systems can be seen in the long-standing economic inequality along racial lines.

Ride-hailing platforms like Uber and Lyft have transformed urban transportation, yet their algorithms risk perpetuating systemic inequities. Research shows that low-income and majority-minority neighborhoods frequently face higher fares and more frequent surge pricing, even when demand mirrors that of wealthier areas. These disparities stem from models that unintentionally amplify historical biases, such as undervaluing demand in marginalized neighborhoods or prioritizing affluent users. In New York City, where public transit gaps and economic inequality intersect, such algorithmic biases compound mobility challenges for vulnerable populations. This study examines these patterns through a data-driven lens, testing interventions to align pricing systems with equity goals while maintaining operational viability.

Data

Our analysis relies on ride-hailing trip data from the New York City Taxi and Limousine Commission (TLC), which records millions of for-hire vehicle trips from services like Uber and Lyft. This dataset includes details such as trip duration, distance, fare breakdowns, pickup and drop-off locations, and vehicle type. Additionally, we integrate demographic data from the NYC Neighborhood Tabulation Areas (NTA), which provides insights into socioeconomic indicators like income levels, poverty rates, and racial composition. By merging these datasets, we aim to uncover potential biases in pricing while accounting for external factors like geography and infrastructure.

Method

As ride-hailing services continue to shape urban mobility, understanding their pricing mechanisms is crucial for ensuring fairness and accessibility. Our research takes a deep dive into the factors influencing ride-hailing fares, questioning whether these platforms unintentionally reinforce socioeconomic and racial disparities. By systematically analyzing trip data, demographic distributions, and pricing models, we probe into how algorithmic decision-making impacts different communities. Through this investigation, we aim to identify patterns of potential bias, assess fairness in fare structures, and explore strategies to mitigate inequities in ride-hailing services.

Pre-processing

The TLC dataset contains 19,663,930 rows and 24 columns

Data Cleaning

Analysis of 4.56 million NYC TLC trips shows strong correlations between distance, duration, and fare (>0.80), while weaker ties to driver compensation (0.56) identify key areas for fare strategy improvements.

Merge Data

Histograms of NYC's Neighborhood Tabulation Areas reveal stark racial/ethnic segregation patterns, with Asian, Black, and Hispanic populations concentrated in distinct pockets while our merged 60-column dataset integrates ride-hailing metrics with demographic data to analyze fare disparities across socioeconomic communities.

Exploratory Data Analysis Insights

Analysis of NYC ride-hailing reveals demand peaks during weekday rush hours with 35% higher requests during rainfall, alongside price elasticity variations across neighborhoods and 20-30% longer wait times in outer boroughs, highlighting opportunities for zone-specific incentives and equity-focused pricing models.

Probing the Ride-Hailing

We began by cleaning our NYC ride-hailing dataset and computing key features, including trip distance, trip length, and a normalized cost metric—cost per mile (calculated as Base Passenger Fare divided by Trip Distance). To capture potential price discrimination, we defined a binary target, high_cost, by labeling rides as high cost if their cost per mile exceeded the median value. Next, we incorporated fairness into our analysis by creating protected attributes based on neighborhood ethnic demographics (Black, Hispanic, White, Asian, and Other). For each ethnicity, we generated binary indicators by comparing the respective percentage to its median value.

Baseline Model

We trained a standard logistic regression model using the features (trip distance, trip length, and cost per mile) to predict the high_cost target.

Equalized Odds

Equalized Odds balances the False Positive and False Negative Rates across protected groups. This ensures that misclassification risks—whether overcharging or undercharging—remain similar for all groups, reducing discriminatory outcomes.

ExponentiatedGradient

ExponentiatedGradient enforces fairness constraints like Equalized Odds by iteratively reweighting training samples to minimize error while satisfying the fairness requirement. This method balances accuracy and fairness, reducing bias without a major performance loss.

Fairness-aware model

We applied ExponentiatedGradient with an Equalized Odds constraint to balance false positive and negative rates across protected groups, thereby mitigating bias.

Result

We evaluated the performance of both a baseline logistic regression model and a fairness-aware model (using the ExponentiatedGradient algorithm with an Equalized Odds constraint) across several protected attributes representing different ethnic groups (Black, Hispanic, White, Asian, and Other). In all cases, the baseline model achieved an accuracy of 87.8%. However, when fairness constraints were applied, overall accuracy generally decreased as a trade-off for reducing bias.