Toward Equitable Low-Carbon Mobility: Fairness-Aware Demand Prediction for Expanding Bike-Sharing Systems

Man Luo; Yixuan Zhao

Source-linked AI summary

Toward Equitable Low-Carbon Mobility: Fairness-Aware Demand Prediction for Expanding Bike-Sharing Systems

Man Luo, Yixuan Zhao

arXiv:2608.26451v1Aug 26, 2026cs.LG

TL;DR

Expanding bike-sharing networks create a cold-start gap because new stations lack historical demand, while historical ridership can encode income-based inequities. FairGIN combines expansion simulation, adaptive knowledge transfer, and fairness-aware optimization to predict new-station demand and support equitable deployment. Across NYC and Seattle, it reports leading accuracy and substantial reductions in income-based disparities, while the evaluated fairness gains vary with expansion conditions and baseline comparisons.

Problem
New stations lack historical ridership, and historical demand may reflect infrastructure inequalities, creating a joint prediction and income-based fairness challenge during network expansion.
Method
FairGIN uses expansion-simulated increment training, attention-based knowledge transfer, and income-stratified fairness training with deployment scoring in a unified dynamic graph framework.
Results
31.2% reduction in RFG relative to the strongest Seattle baseline, alongside a 33.3% reduction on NYC, demonstrates substantial income-based disparity reductions across both cities.
Takeaways & Limitations
Fairness-aware design at the embedding level can support more inclusive low-carbon mobility access while retaining leading prediction accuracy in the evaluated urban settings.
Takeaways & Limitations
FairGIN’s robustness advantage depends on expansion conditions, with stability becoming more pronounced across varying pseudo-new station sizes and compositions and fairness remaining effective at r = 5%.

Abstract

from arXiv · show

Bike-sharing systems are an important component of low-carbon urban mobility, but continued expansion creates challenges in both cold-start prediction and equitable resource allocation. Newly deployed stations lack historical ridership records, causing a mismatch between training and inference for graph-based models on evolving networks. Historical demand may also encode structural inequalities, as lower ridership in low-income neighborhoods can reflect limited infrastructure access rather than weak latent demand. Models trained directly on such data may therefore reinforce existing mobility disparities. We propose FairGIN, a fairness-aware graph neural network for demand prediction in expanding bike-sharing systems. FairGIN integrates three components. Expansion-Simulated Increment Training stochastically simulates network expansion during training to reduce the cold-start distribution gap. Attention-Based Knowledge Transfer combines station-adaptive temperature scaling with orthogonal embedding alignment to transfer representations from data-rich existing stations to data-sparse new stations. Fairness-Aware Optimization introduces income-stratified regularization and an equity-calibrated deployment score to support more inclusive station placement. Experiments on NYC and Seattle demonstrate that FairGIN achieves state-of-the-art predictive accuracy across diverse expansion scenarios while substantially reducing income-based disparities without compromising overall system efficiency.

I. INTRODUCTION

Bike-sharing expansion creates coupled cold-start and equity challenges: new stations lack demand histories, while historical ridership and uneven coverage can encode income-based disparities. FairGIN addresses these challenges with expansion simulation, adaptive knowledge transfer, fairness regularization, and validation across NYC and Seattle.

Motivation: 22.8 versus 6.3 stations per 10,000 residents shows that Manhattan’s highest-income quintile has more than three times the coverage of its lowest-income quintile.The highest-income quintile exceeds $160k, while the lowest-income quintile is below $55k.
Equity Challenge: Observed low ridership in underserved communities may reflect affordability constraints, limited technology adoption, and absent nearby stations rather than weak latent demand.Using such observations uncritically can contribute to self-reinforcing underestimation of community need.
Problem: New stations lack historical ridership and appear in continuously evolving spatial networks, complicating accurate demand prediction for fair resource allocation.Traditional regression models capture temporal regularities but have limited capacity to model inter-station spatial dependencies.
Proposed Approach: FairGIN combines expansion-simulated increment training, attention-based knowledge transfer, and income-stratified fairness regularization for expanding bike-sharing networks.The components target the training-inference graph gap, cold-start representation learning, and systematic prediction disparities, respectively.
Evaluation: NYC and Seattle experiments demonstrate superior performance compared with state-of-the-art baselines across the evaluated mobility systems.The supplied introduction reports consistent experimental superiority without providing metric values here.

II. PROBLEM STATEMENT

FairGIN formulates new-station demand prediction on an expanding dynamic graph using existing-station features, new-station spatial features, and graph structure. Because new stations lack historical demand, the model must infer their hourly demand from spatial context and transferred knowledge.

Expanding Network: The expanding system contains existing stations V_A and newly deployed stations V_B, represented within a dynamic graph G_t.Existing and newly deployed stations are modeled as distinct station sets whose union forms the network vertex set.
Station Features: Location-based features are available for both existing and new stations and include points of interest, roads, weather, and ambient taxi flow.Ambient taxi flow serves as a proxy for surrounding mobility demand.
Station Features: Historical hourly demand sequences are available only for existing stations; newly deployed stations have no ridership history.For new stations, the history-based feature vector is set to zero in the formulation.
Graph Construction: Graph edge weights reflect station availability using spatial similarity, geographic distance decay, and temporal demand correlation when historical observations exist.The temporal term is omitted for edges involving newly deployed stations because their demand histories are unavailable.
Prediction Task: The task learns a mapping f_θ from existing-station features, new-station spatial features, and graph G to hourly demand predictions for new stations.New stations are unseen during training, so their demand is inferred from spatial context and knowledge transferred from existing stations.

C. Fairness Formulation

FairGIN formulates fairness-aware prediction for newly deployed stations by measuring income-based disparities and jointly optimizing accuracy with fairness. Its station representations combine spatial, temporal, and graph-neighborhood information despite unequal feature availability.

C. Fairness Formulation: Candidate stations are divided into advantaged G+ and disadvantaged G− groups using city-wide median household income at the census-tract level.Stations in tracts at or above the median form G+; those below form G−.
C. Fairness Formulation: Region-based Fairness Gap measures the per-capita predicted-demand difference between income groups across evaluated regions.A smaller RFG indicates lower group-level prediction disparity.
C. Fairness Formulation: Individual-based Fairness Gap measures station-level disparity by representing each station as a weighted mixture of income groups.The paper also reports Spearman’s ρ between predicted demand and tract-level income, where ρ ≈0 indicates weak income dependence.
C. Fairness Formulation: FairGIN jointly minimizes prediction error and income-based demand disparity through observation, simulation, and fairness losses.Lobs supervises existing stations, Lsim supervises expansion-simulated pseudo-new stations, and Lfair penalizes per-capita disparity between G+ and G−.
C. Fairness Formulation: During inference, candidate stations receive an income-aware deployment score combining predicted demand with an equity bonus for G−.The bonus α ≥0 controls deployment prioritization for disadvantaged-group stations.
III. METHODOLOGY: FairGIN is designed to jointly address evolving station-graph topology and equity risks in data-driven deployment decisions.The framework targets fairness-aware demand prediction during network expansion.
A. Station Representation and Graph Construction: Existing-station embeddings combine spatial features and temporal demand sequences, whereas new-station embeddings use spatial features only because historical observations are unavailable.Spatial inputs include POI distributions, road-network characteristics, and taxi flow; temporal sequences are encoded by a GRU.
A. Station Representation and Graph Construction: Dynamic graph convolution propagates contextual information using normalized adjacency and learnable transformations, producing representations that encode station features and neighborhood structure.These representations support demand prediction for new stations with limited observations.

B. Expansion-Simulated Increment Training

Expansion-Simulated Increment Training reduces the mismatch between fully observed training stations and newly deployed stations without historical demand. It simulates network expansion while retaining supervision and income labels for sampled pseudo-new stations.

B. Expansion-Simulated Increment Training: Training only on fully observed stations creates a feature-distribution mismatch because deployment includes nodes lacking temporal histories.This mismatch can degrade prediction performance for newly deployed stations.
B. Expansion-Simulated Increment Training: Expansion-Simulated Increment Training simulates network expansion during every training epoch.The strategy reproduces the cold-start condition while preserving location-based features.
B. Expansion-Simulated Increment Training: 15% of existing stations are uniformly sampled as pseudo-new stations during each epoch.The sampled stations are denoted V′B.
B. Expansion-Simulated Increment Training: The remaining stations form the observed set in a graph where observed and history-limited stations coexist.The resulting structural setting matches the deployment graph.
B. Expansion-Simulated Increment Training: Pseudo-new stations retain ground-truth demand for simulation supervision and geographic coordinates for census-tract income-group assignment.This enables fairness-loss optimization without additional annotations.

C. Attention-Based Knowledge Transfer

Attention-Based Knowledge Transfer makes new-station representations prediction-ready by selectively aggregating information from observed stations. Station-specific temperature scaling, orthogonal alignment, and gated fusion balance transferred knowledge with local spatial information.

C. Attention-Based Knowledge Transfer: FairGIN aggregates knowledge from observed existing stations according to representation similarity for new-station prediction.A differentiable transfer mechanism addresses the absence of historical demand information.
C. Attention-Based Knowledge Transfer: Similarity scores between each new station and observed existing stations determine the attention distribution.The scores are computed from the corresponding station embeddings.
C. Attention-Based Knowledge Transfer: A station-specific temperature is learned from each new-station embedding to control attention concentration.Smaller τj emphasizes the most similar stations, whereas larger τj encourages broader aggregation.
C. Attention-Based Knowledge Transfer: The transferred representation is orthogonally transformed to reduce distributional mismatch between existing and newly deployed station embeddings.The mapping preserves orthogonality during training through a Cayley-transform parameterization.
C. Attention-Based Knowledge Transfer: A learnable gate combines the transformed transfer signal with each station’s local spatial representation.This allows the model to balance transferred knowledge and local information according to station context.
C. Attention-Based Knowledge Transfer: A shared MLP decoder converts the resulting representation into demand predictions.The decoder is applied after transfer and gated fusion.

D. Fairness-Aware Optimization

Fairness-Aware Optimization combines accuracy supervision, pseudo-new-station simulation supervision, and income-based fairness penalties. At deployment, an equity-calibrated score adjusts demand rankings without changing the trained model.

D. Fairness-Aware Optimization: FairGIN training uses observation, simulation, and fairness loss terms in a joint objective.The observation loss preserves accuracy on existing stations, while the other terms address pseudo-new stations and income-based disparity.
D. Fairness-Aware Optimization: The fairness loss penalizes income-based per-capita prediction disparity among pseudo-new stations.The hyperparameters λsim and λfair are selected by validation-set grid search.
D. Fairness-Aware Optimization: During inference, ESIT is disabled and FairGIN predicts candidate-station demand using deployment-time spatial features without historical observations.The attention-gate pipeline aggregates relevant existing-station knowledge before decoding demand.
D. Fairness-Aware Optimization: FairGIN can be applied directly to newly deployed stations without retraining, graph reconstruction, or manual feature imputation.This follows from aligning ESIT training with the inference condition.
D. Fairness-Aware Optimization: The deployment score adds an equity bonus to predicted demand for stations in G−.The bonus is calibrated at the decision level, and experiments compare demand-only ranking with fairness-aware strategies.

IV. EXPERIMENTS

The experiments evaluate FairGIN on NYC and Seattle bike-sharing data using common new-station protocols, accuracy and fairness metrics, and comparisons with nine baseline methods.

Datasets: Experiments use publicly available NYC and Seattle bike-sharing datasets with income variation across urban neighborhoods.NYC contains 746 existing and 312 new stations; Seattle contains 186 existing and 61 new stations.
Implementation: FairGIN uses spatial and temporal encoders with graph convolutional layers and is implemented in PyTorch Geometric.The spatial encoder is a two-layer MLP, the temporal encoder is a single-layer GRU, and the model uses L = 2 graph convolutional layers.
Compared methods: FairGIN is compared with nine baselines spanning statistical, sequential, graph-based, inductive, fairness-aware, and urban foundation models.The compared methods include ARIMA, LSTM, STGCN, DCRNN, FairST, GraphSAGE, DA-MRGNN, KITS, and UrbanGPT.
Compared methods: Transductive baselines receive zero temporal features for new stations, whereas inductive methods use spatial features or knowledge transfer to support unseen-node prediction.This protocol reflects the absence of historical demand observations at newly deployed stations.
Evaluation protocol: Evaluation measures new-station prediction accuracy with MAE and RMSE, and fairness with RFG, IFG, and Spearman |ρ|.All methods are trained on existing stations and evaluated on new stations under the same protocol.

B. Main Performance Comparison

FairGIN consistently combines strong new-station accuracy with lower income-related disparities across NYC and Seattle. Its components support performance under ablations, varying fairness settings, expansion patterns, and expansion rates.

Main comparison: FairGIN achieves the lowest MAE and RMSE and the smallest RFG, IFG, and Spearman |ρ| on both NYC and Seattle.This joint result indicates that accuracy gains are not achieved at the expense of fairness across income groups.
Main comparison: 15.8% lower MAE and 15.3% lower RMSE than KITS are reported on NYC, while RFG and IFG decrease 33.3% and 37.3% versus FairST.KITS is the strongest accuracy baseline and FairST is the strongest fairness baseline in the reported comparisons.
Main comparison: Fixed-topology and statistical methods struggle with cold-start prediction because new stations lack historical demand and the inference graph differs from training.Inductive and cold-start-oriented methods improve accuracy, but their fairness gains remain limited without income-stratified supervision or explicit disparity constraints.
Ablation study: Removing ESIT, knowledge transfer, fairness learning, or temperature scaling causes clear performance degradation on both datasets.The ablation study evaluates w/o ESIT, w/o KT, w/o FL, and w/o TS variants against the full framework.
Sensitivity analysis: The best ESIT mask ratio is ρ = 0.15, balancing simulated expansion diversity against observed-graph completeness.Performance improves from ρ = 0.05 to 0.15 and declines above 0.15 as message passing and the knowledge-transfer reference pool weaken.
Robustness analysis: FairGIN remains the lowest-MAE and lowest-RFG method across expansion rates from 5% to 30%, with changes of at most 5.6% in MAE and 11.3% in RFG.The expansion-rate analysis uses five independent trials for each setting.

E. Computational Complexity Analysis

FairGIN combines cold-start prediction and income-aware fairness within a moderately sized model, while its training and inference costs remain competitive with deep-learning baselines.

Model complexity: FairGIN has 211.2K trainable parameters, comparable to LSTM and DA-MRGNN and moderately larger than KITS.The added capacity supports cold-start prediction and income-aware fairness optimization in one model.
Runtime: FairGIN trains approximately eight times faster than DCRNN, whose recurrent graph architecture unrolls T = 24 graph-convolution steps.FairGIN encodes historical demand with a GRU before graph propagation, limiting additional transformation and fusion overhead.
Deployment equity: The deployment simulation measures the selected disadvantaged-group share and mean predicted demand under top-10, top-20, and top-30 thresholds.Demand-only selection is compared with random selection and FairGIN’s equity-aware scoring.
Deployment equity: FairGIN raises the G−ratio to 0.60 for NYC top-10 selections and 0.58 for Seattle, while increasing selected G−station demand from 2.31 to 2.47 on NYC and 2.04 to 2.19 on Seattle versus random selection.The reported gains improve representation without indiscriminately selecting low-demand locations.

G. Equity-Aware Deployment Case Study

The case study shows how FairGIN’s equity-calibrated deployment score changes candidate rankings toward low-income stations, while experiments also examine fairness, robustness, and limitations.

Equity-Aware Deployment: Four high-income stations are replaced by four low-income stations after adding 0.2 to disadvantaged-group scores.The displaced high-income scores range from −0.038 to 0.007, while replacement scores range from 0.017 to 0.130.
Equity-Aware Deployment: The adjustment mainly changes rankings among candidates with relatively similar predicted demand rather than promoting consistently low-potential stations.Original demand differences between replaced and elevated stations range from 0.06 to 0.19 normalized units.
Budget Sensitivity: With α = 0.2, FairGIN selects only low-income-group stations for every budget k≤10, whereas demand-only ranking selects 60% at k = 5 and k = 10.Low-income candidates constitute 0.625 of the candidate set.
Robustness: At 30% expansion, FairGIN’s MAE variation across random seeds remains within 5.6%, and its stability advantage over baselines becomes more pronounced.At 5% expansion, FairGIN retains a clear RFG advantage over KITS.
Limitations: FairGIN’s binary median-income grouping masks meaningful error variation within the disadvantaged group.MAE rises from 0.588 for the near-median subgroup to 0.717 for the low-income subgroup and 1.586 for stations without ACS records.
Limitations: The study uses income as its only protected attribute, static spatial features, and U.S. docked systems with relatively dense ACS coverage.Generalization to dockless networks, different income–mobility relationships, and data-scarce Global South regions remains unexamined.

VI. RELATED WORK

Related work spans sustainable urban mobility, bike-sharing expansion, fairness-aware transportation modeling, dynamic graph prediction, and cold-start knowledge transfer.

Research Context: The literature review situates FairGIN’s attention-based knowledge transfer within broader work on fairness, dynamic graphs, and cold-start prediction.It explicitly identifies knowledge transfer and cold-start techniques as relevant strands of prior research.
Sustainable Urban Mobility: Bike-sharing is positioned as a flexible, zero-emission mobility option that can reduce private-vehicle reliance and support urban decarbonization.The passage frames bike-sharing as part of sustainable urban transport infrastructure.

B. Bike-Sharing System Expansion

Prior bike-sharing expansion research commonly uses historical demand, spatial heuristics, or coarse zones, while FairGIN models dynamic network evolution, fine-grained demand, and equity together.

Expansion Planning: Existing expansion studies target new-station deployment and capacity adjustment using spatial clustering, demand mapping, optimization, historical usage, or static heuristics.These approaches guide expansion decisions but are not described as jointly modeling evolving station interactions and fairness.
Demand Prediction: Zone-based approaches estimate average demand at newly deployed stations, whereas FairGIN models dynamic network evolution and changing spatial-temporal station interactions.FairGIN also predicts instantaneous and expected demand at fine-grained temporal resolution.
Fairness: Fairness-aware machine learning provides group- and individual-fairness paradigms relevant to evaluating disparities in station demand prediction.Group fairness includes demographic parity and equalized odds, while individual fairness concerns similar predictions for similarly situated instances.
Dynamic Graph Prediction: Dynamic graph methods model transportation dependencies with recurrent, convolutional, or graph-based architectures but face computational, stability, or oversmoothing concerns.The cited review describes recurrent approaches as costly and potentially unstable over long sequences, while repeated graph propagation may oversmooth fine-grained patterns.
Cold-Start Learning: Knowledge transfer addresses cold-start data scarcity by using representations from data-rich sources, with GraphSAGE enabling embeddings for unseen nodes.Cold Brew extends this direction to strict feature-only cold-start inference.
Positioning FairGIN: FairGIN combines expansion simulation, attention-based transfer, and fairness-aware training with deployment scoring in an end-to-end framework.Experiments on NYC Citi Bike and Seattle Bikeshare report leading accuracy alongside reduced income-group demand disparities.