Source-linked AI summary
Learning to Transfer Across Modes: Towards Unified Urban Mobility Forecasting
Yixuan Zhao, Man Luo
TL;DR
Urban mobility forecasting must handle modes with heterogeneous spatial structures and uneven temporal histories. TransMod builds a shared zone-level representation, aligns modes, and transfers spatio-temporal patterns from data-rich sources to data-scarce targets. Experiments report consistent baseline improvements and robustness under several challenging settings.
Problem
Existing forecasting methods largely assume compatible spatial structures and abundant mode-specific histories, leaving joint forecasting across heterogeneous mobility modes insufficiently addressed.
Method
TransMod softly maps station-level information into shared zones, aligns cross-modal representations, and retrieves transferable spatio-temporal patterns through memory-based transfer.
Results
TransMod outperforms state-of-the-art baselines and remains robust under spatial coarsening, data scarcity, cross-city transfer, and long-horizon forecasting.
Takeaways & Limitations
Cross-modal spatio-temporal knowledge transfer is effective for forecasting in data-heterogeneous urban mobility settings.
Takeaways & Limitations
Future work should test TransMod with additional mobility systems and improve adaptability through online or continual memory updates.
Abstract
from arXiv · showhide
Urban transportation systems consist of multiple mobility modes that coexist within the same city and exhibit complex interdependencies, leading to correlated demand dynamics across modes. However, forecasting demand jointly across different modes remains challenging due to substantial heterogeneity in space and the limited availability of historical data for emerging modes. Existing forecasting methods are largely developed for individual mobility modes and implicitly assume compatible spatial structures between source and target systems, which severely restricts their applicability in multi-modal settings. To address these challenges, we propose \textbf{TransMod}, a unified framework for urban mobility demand forecasting that enables effective knowledge transfer across heterogeneous mobility modes. TransMod constructs a shared zone-level spatial representation that aligns mobility systems with different spatial granularities into a common space, thereby reducing structural mismatch and distributional shift. Built on this unified representation, TransMod further learns transferable spatio-temporal patterns from data-rich source modes and adapts them to data-scarce target modes, alleviating the dependence on extensive target-domain histories. Extensive experiments on real-world datasets demonstrate that TransMod consistently outperforms existing approaches and provides robust forecasting performance under limited target data.
1 Introduction
Urban mobility modes share correlated demand dynamics but differ substantially in spatial representation and temporal data availability, making joint forecasting difficult. TransMod addresses this gap by creating a unified spatial foundation and transferring knowledge from data-rich modes to data-scarce ones.
- Motivation: Multiple mobility modes share underlying demand drivers, producing correlated and temporally synchronized dynamics despite differences in service mechanisms and infrastructure.Commuting peaks, weather changes, and major urban events can affect multiple services simultaneously, although responses vary by spatial scale, service type, and usage context.
- Challenges: Station-based systems offer fine-grained spatial observations and rich histories, whereas emerging station-less services require coarser aggregation and often have limited, noisy observations.Bike-sharing and metro use fixed station anchors, while ride-hailing records spatially dispersed origin–destination flows.
- Research Gap: Existing forecasting and transfer methods largely target individual systems or assume comparable data structures, limiting their use across heterogeneous modes with different spatial and temporal scales.These assumptions are restrictive when multiple interacting modes coexist within the same city.
- TransMod: TransMod formulates cross-modal forecasting in a shared zone-level space by softly aggregating station-level information while preserving local spatial proximity.The framework uses a learnable, geography-aware assignment to establish a common spatial foundation across mobility modes.
- TransMod: TransMod aligns mobility-mode representations and transfers learned spatio-temporal dynamics from source systems with longer histories to target systems with limited temporal data.Its transfer design is intended to forecast target demand without requiring historical observations from the target system.
- Evidence: Experiments on real-world shared mobility datasets report that TransMod consistently outperforms strong baselines and achieves state-of-the-art zone-level demand forecasting.The reported evaluation supports the framework’s effectiveness across heterogeneous mobility settings.
2 Preliminaries
The paper models heterogeneous mobility systems using modality-specific station graphs and a shared zone graph. It formulates forecasting as predicting zone-level ride-hailing demand from historical source-mode observations and static target-mode spatial features.
- Problem Setting: The setting includes station-based bike-sharing and metro systems alongside zone-based ride-hailing, with mismatched spatial granularity and temporal availability.Station-based demand is recorded at fixed geographic anchors, while ride-hailing demand is aggregated over spatial zones.
- Station Graph: For each station-based modality, contextual features encode location, POIs, road-network characteristics, and weather conditions.These features supplement station demand observations with spatial and environmental context.
- Station Graph: Dynamic station graphs represent spatial and behavioral dependencies using geographic proximity and time-varying origin–destination interactions.The corresponding weighted adjacency matrix changes with time through these dependency patterns.
- Zone Graph: The city is partitioned into non-overlapping 1 km×1 km grid zones, and a zone graph models geographic adjacency and road-network connectivity.Ride-hailing demand is naturally observed at this zone level.
- Problem Formulation: Given historical demand and graph observations from bike-sharing and metro, plus auxiliary features and static ride-hailing spatial features, the objective is to forecast zone-level ride-hailing demand.Historical ride-hailing temporal observations are unavailable in this formulation.
3 Methodology
TransMod unifies heterogeneous mobility modes in a shared zone-level space, then transfers spatio-temporal knowledge from source systems to target forecasting under sparse observations.
- Framework overview: TransMod integrates station-level spatio-temporal representation learning, zone-level spatial unification, and cross-modal knowledge transfer.The framework is designed for heterogeneous mobility systems with sparse target observations.
- Station-zone representation aggregation: Soft station-to-zone assignments aggregate fine-grained source information into spatially continuous zone representations while preserving geographic locality.Assignments are normalized, initialized from Gaussian distance kernels, restricted to nearby zones, and refined using observed data with geographic regularization.
- Unified zone graph alignment: Aligned bike-sharing and metro zone representations are fused into unified embeddings and a graph combining mobility interactions, geographic proximity, and feature similarity.Global distribution alignment uses MMD, while InfoNCE encourages same-zone representations from different modes to be close.
- Knowledge transfer across modes: A memory mechanism retrieves transferable temporal prototypes for ride-hailing from queries combining aligned source representations with temporal context.The prototypes index joint spatial-temporal configurations, while diversity regularization encourages distinct patterns.
- Training objective: Source pre-training jointly optimizes forecasting and representation objectives for prediction, cross-modal alignment, geographic assignment regularization, and prototype diversity.The composite objective balances these components through corresponding loss weights.
4 Experiments
The experiments evaluate TransMod on real-world multi-modal mobility datasets from New York City and Chicago, using bike-sharing, metro, and ride-hailing data with auxiliary urban features.
- Evaluation design: The evaluation includes comparisons with baseline models, computing-efficiency analysis, ablations, and robustness analysis across scenarios.These experiments are designed to assess TransMod’s performance under different evaluation settings.
- Datasets: Datasets from New York City and Chicago combine bike-sharing, metro, and ride-hailing records with points of interest, weather, and road-network information.The evaluation integrates multiple mobility modes and auxiliary features.
- Data access protocol: The stage-wise data access protocol distinguishes used, unavailable, and frozen data during the experimental procedure.The table defines ✓ as used, × as unavailable, and frz as frozen and not re-accessed.
- Data modalities: Bike-sharing and metro data provide station-level demand records, while ride-hailing trips are aggregated into predefined spatial zones.The collection includes NYC Citi Bike, Chicago Divvy, NYC MTA, Chicago CTA, NYC TLC, and Chicago TNP Trips data.
4.3 Evaluation Metrics and Compared Methods
TransMod is evaluated with RMSE, MAE, and MAPE against single-system, spatio-temporal, transfer, and cross-modal baselines on transfer tasks across mobility modes. It consistently achieves the lowest reported errors across cities and metrics.
- Evaluation Metrics and Compared Methods: RMSE, MAE, and MAPE provide complementary measures of forecasting error and demand-variation capture.The compared methods span classical time-series, recurrent, spatio-temporal, transfer-learning, and cross-modal approaches.
- Evaluation Metrics and Compared Methods: Traditional and non-transfer spatio-temporal models perform worse because they lack cross-modal information or cannot resolve station–zone structural mismatch.Transfer and cross-modal methods improve accuracy but still struggle with spatial granularity discrepancies and temporal sparsity.
- Evaluation Metrics and Compared Methods: TransMod consistently outperforms all competing methods across cities and evaluation metrics.Table 2 marks the best results in bold, while the reported comparison states that TransMod achieves the lowest MAE, RMSE, and MAPE.
4.5 Computing Efficiency Analysis
TransMod has graph-based pre-training costs comparable to competing transfer methods but substantially reduces adaptation time and maintains competitive inference latency. Its efficient adaptation freezes the memory and spatial mapping modules while fine-tuning only lightweight components.
- Computing Efficiency Analysis: Pre-training cost is comparable to other graph-based transfer methods because TransMod performs cross-modal representation alignment and memory learning.The efficiency evaluation reports pre-training time, adaptation time, and per-sample inference latency.
- Computing Efficiency Analysis: Sub-10 ms inference latency results from memory lookup followed by zone-level propagation instead of recomputing full station-level graph attention.Memory retrieval has complexity O(N_z·M·d), with M=128 fixed regardless of city scale, so inference cost grows linearly with the number of zones.
4.6 Ablation Studies
Ablation studies show that TransMod’s soft assignment mechanism is most critical, while uncertainty embeddings, cross-modal alignment, and memory-based transfer also contribute to performance. Removing these components degrades forecasting accuracy on NYC and Chicago.
- Ablation Studies: Removing soft assignment increases MAE by 13.0% in NYC and 16.0% in Chicago, the largest degradation among variants.Hard nearest-zone assignment can distort station-to-zone aggregation, whereas geography-regularized soft assignment bridges spatial granularity mismatch.
- Ablation Studies: Replacing memory-based transfer with a GRU decoder increases MAE by 8.8% in NYC and 11.8% in Chicago.The memory pool captures richer zone-structure-conditioned temporal patterns than the GRU decoder in this comparison.
- Ablation Studies: Removing Gaussian uncertainty embeddings or cross-modal alignment causes moderate performance drops.Uncertainty modeling suppresses unreliable auxiliary signals, while alignment keeps source representations distributionally consistent for memory learning.
4.7 Robustness Analysis
TransMod is evaluated across cross-city and cross-modal transfer, practical perturbations, and auxiliary-feature degradation, with results indicating robust forecasting under data-limited heterogeneous conditions.
- Cross-city transfer: With only 10% of Chicago data, TransMod achieves an RMSE of 12.38, compared with 16.21 without transfer.The transferred model approaches the full-data upper bound as target-city data become limited.
- Cross-modal transfer: TransMod achieves the lowest MAPE across all six source–target configurations for ride-hailing and bike-sharing prediction.The evaluation uses bike-sharing, metro, or ride-hailing as auxiliary sources in both target-task directions.
- Practical perturbations: Under spatial coarsening from 0.5 km to 2 km, TransMod’s RMSE rises by only 9.5%, compared with 17%–24% for competing baselines.The analysis also reports consistent superiority under data scarcity and comparable performance using only 30% of ride-hailing records.
- Auxiliary feature degradation: Removing POI features increases MAE by 5.8% in NYC and 6.1% in Chicago, the largest degradation among individual auxiliary-feature categories.Removing all auxiliary features increases MAE by 10.8% in NYC and 11.8% in Chicago, while degradation remains noncatastrophic.
4.8 Hyperparameter Sensitivity Analysis
The sensitivity analysis varies memory-pool size, InfoNCE temperature, and bike-sharing bandwidth across NYC and Chicago, highlighting bandwidth behavior and its spatial-assignment implications.
- Analysis setup: TransMod sensitivity experiments vary memory pool size M, InfoNCE temperature τ, and bike-sharing bandwidth σ_bike while reporting MAE on NYC and Chicago.Default settings are M=128, τ=0.07, and σ_bike=500 m.
- Bandwidth sensitivity: Reducing σ_bike to 300 m raises NYC MAE from 10.78 to 11.34 by making assignment overly localized and closer to hard nearest-zone mapping.Increasing bandwidth to 800 m causes a milder degradation, with NYC MAE increasing to 11.05.
- Bandwidth sensitivity: Stable performance between 500 m and 800 m indicates that the soft assignment mechanism is robust to moderate bandwidth variation.The passage attributes the larger degradation at 300 m to overly localized assignment and the degradation at 800 m to distant-zone noise.
5 Related Work
Related work covers urban demand forecasting, spatio-temporal representation learning, and transfer learning, while identifying persistent incompatibility across heterogeneous mobility systems.
- Urban Mobility Demand Forecasting: Urban mobility demand forecasting models temporal dynamics and spatial correlations, but most methods assume internally consistent spatial units, observations, and demand-generation processes within one system.This restricts direct transfer when systems represent shared demand through stations, stops, or zones.
- Spatio-Temporal Representation Learning: Graph-based models such as DCRNN, STGCN, and Graph WaveNet represent sensors, stations, or regions as nodes and spatial proximity or mobility interactions as edges.Adaptive graph structures and node embeddings improve flexibility for evolving intra-system spatial correlations.
- Spatio-Temporal Representation Learning: Despite adaptive and attention-based advances, most representation-learning methods remain tied to fixed spatial units and compatible spatial semantics across source and target observations.This limits applicability when node sets and observation structures differ between systems.
- Transfer Learning and Domain Adaptation: Transfer-learning approaches reuse spatio-temporal patterns from data-rich source domains to improve forecasting in data-scarce target domains.Existing approaches commonly use distribution matching, adversarial learning, or feature-level invariance.
6 Conclusion and Future Work
The conclusion presents TransMod as a unified approach for cross-modal forecasting with mismatched spatial granularity and uneven temporal observability, while outlining broader extensions.
- Conclusion: TransMod builds shared spatial representations, aligns cross-modal representations, and retrieves transferable temporal patterns through memory-based transfer.Experiments on New York City and Chicago report robustness under spatial coarsening, data scarcity, cross-city transfer, and long-horizon forecasting.
- Conclusion: The reported results highlight cross-modal spatio-temporal knowledge transfer as effective in data-heterogeneous urban settings.The paper frames this as a scalable step toward generalizable urban mobility forecasting.
- Future Work: Future work proposes adding shared scooters, on-demand buses, and autonomous fleets to test generality under broader urban heterogeneity.The proposed extensions would broaden the mobility systems evaluated by TransMod.
- Future Work: Online or continual memory updates may improve adaptability to evolving mobility patterns and sudden demand shifts.Multi-city pre-training and more interpretable memory prototypes are also proposed.