MoDE-Boost: Boosting Shared Mobility Demand with Edge-Ready Prediction Models

Antonios Tziorvas; George S. Theodoropoulos; Yannis Theodoridis

cs.LGFeb 18, 2026

MoDE-Boost: Boosting Shared Mobility Demand with Edge-Ready Prediction Models

Authors: Antonios Tziorvas, George S. Theodoropoulos, Yannis Theodoridis

Organizations: Department of Informatics, University of Piraeus, Piraeus, Greece.

Abstract

Urban demand forecasting plays a critical role in optimizing routing, dispatching, and congestion management within Intelligent Transportation Systems. By leveraging data fusion and analytics techniques, traffic demand forecasting serves as a key intermediate measure for identifying emerging spatial and temporal demand patterns. In this paper, we tackle this challenge by proposing two gradient boosting model variations, one for classiffication and one for regression, both capable of generating demand forecasts at various temporal horizons, from 5 minutes up to one hour. Our overall approach effectively integrates temporal and contextual features, enabling accurate predictions that are essential for improving the efficiency of shared (micro-) mobility services. To evaluate its effectiveness, we utilize open shared mobility data derived from e-scooter and e-bike networks in five metropolitan areas. These real-world datasets allow us to compare our approach with state-of-the-art methods as well as a Generative AI-based model, demonstrating its effectiveness in capturing the complexities of modern urban mobility. Ultimately, our methodology offers novel insights on urban micro-mobility management, helping to tackle the challenges arising from rapid urbanization and thus, contributing to more sustainable, efficient, and livable cities.

Figures & tables

Figure 1: Impact of forecast‑driven fleet rebalancing in Rotterdam. Left panel: pre‑rebalancing fleet with red circles (demand $>$ supply) and blue circles (demand $<$ supply) indicating shortages and surpluses, respectively. Right panel: post‑rebalancing after moving scooters (orange curved arrows) – shortage areas turn light orange, surplus circle fade to teal

Figure 2: Two indicative timeseries for two regions of Rotterdam: Overschie (top) and Rotterdam Centrum (bottom). The x-axis denotes time, while the y-axis indicates the measured demand values in number of shared vehicles.

Figure 3: Daily demand across different districts in the Rotterdam dataset. Each row represents a district, and each column corresponds to a day in the observation period. Color intensity reflects the total daily demand (aggregated from minute-level data), with warmer colors indicating higher activity.

Figure 4: Average hourly demand across the week in Rotterdam Centrum. The coloring indicates the relative demand intensity and, along with the peaks, aids in highlighting daily and weekly usage patterns.

Figure 5: Temporal evolution of the shared mobility demand (in hourly basis) in Rotterdam Centrum. Seasons are color-coded to highlight the different behaviors. During fall and winter seasons, demand patterns are more stable, whereas during spring and summer seasons, short-term spikes are observed often at “random” periods.

Figure 6: Overview of our proposed SMDF framework: (1) a preprocessing stage creates enriched per‑region timeseries; (2) postprocessing normalizes and encodes features while preserving inter‑region demand ratios; (3) an XGBoost predictor produces the demand forecasts.

Category	Type	Feature Name	Description
Base	ID	region (station)	Spatial entity identifier
Base	value	demand	Raw demand value
Temporal	Lagged	demand_lag_*	Lagged demand values $n$ minutes ago; $n\in\{1,5,60,1440\}$
	Rolling	demand_rolling_mean_*	Rolling mean of demand over past $n$ time steps; $n\in\{5,10,60,1440\}$
	Rolling	demand_rolling_max_*	Rolling maximum of demand over past $n$ time steps; $n\in\{5,10,60,1440\}$
	EWMA	demand_ewm_*	Exponentially weighted moving average of the demand value over past $n$ time steps; $n\in\{5,10,60,1440\}$

Table 1: Summary table of the extracted features used in our approach

Variation	Dataset	Time Span	# Records	# Spatial Entities (regions or points)
Region-based	Rotterdam	2024-08-11 - 2025-07-18	7,561,051	13
	Amsterdam	2024-11-11 - 2025-07-18	32,925,151	88
	The Hague	2024-12-11 - 2025-07-18	13,099,811	46
Point-based	CitiNYC	2021-01-01 - 2022-06-01	36,323,801	1374
Point-based	Divvy	2021-01-01 - 2022-06-01	7,104,591	1007

Table 2: Summary table of the datasets used in our experimental study.

Dataset	Model	Horizon (min)	Train Size (records)	Training Time (sec) (mean ± std)	Test Size (records)	Inference Time (µsec)
Amsterdam	Regressor	5	24.8M	29.83 ± 7.54	2.7M	0.34
		15	24.8M	8.59 ± 1.19	2.7M	0.32
		30	24.7M	15.15 ± 2.73	2.7M	0.32
		60	24.7M	13.63 ± 1.97	2.7M	0.33
	Classifier	5	24.8M	34.43 ± 0.55	2.7M	0.35
		15	24.8M	48.86 ± 0.34	2.7M	0.35

Table 3: Training and inference times of the MoDE-Boost model per dataset, model, and forecasting horizon. The sizes of the training and test sets of the datasets are provided for reference of the training and inference times, respectively.

Dataset	5 min		15 min		30 min		60 min
Dataset	MAE	RMSE	MAE	RMSE	MAE	RMSE	MAE	RMSE
Amsterdam	0.95	1.37	1.28	1.78	1.59	2.12	2.02	2.66
The Hague	1.05	1.52	1.25	1.78	1.44	2.09	1.85	2.61
Rotterdam	4.70	6.33	5.74	7.80	7.14	9.80	9.15	12.05

Table 4: Performance metrics of r-MoDE-Boost across four forecasting horizons.

Dataset	5 min		15 min		30 min		60 min
Dataset	F1	Accuracy	F1	Accuracy	F1	Accuracy	F1	Accuracy
Amsterdam	0.88	0.86	0.85	0.84	0.83	0.81	0.79	0.76
The Hague	0.90	0.84	0.88	0.81	0.87	0.79	0.85	0.76
Rotterdam	0.95	0.90	0.93	0.84	0.92	0.82	0.89	0.77

Table 5: Performance metrics of c-MoDE-Boost across four forecasting horizons.

Dataset	HA		SN		SES		Croston		r-MoDE-Boost
Dataset	MAE	RMSE	MAE	RMSE	MAE	RMSE	MAE	RMSE	MAE	RMSE
Amsterdam	5.88 ( $\pm$ 3.04)	7.21 ( $\pm$ 3.66)	5.35 ( $\pm$ 2.35)	6.86 ( $\pm$ 3.07)	2.29 ( $\pm$ 0.98)	3.08 ( $\pm$ 1.28)	2.29 ( $\pm$ 0.98)	3.08 ( $\pm$ 1.28)	2.06 ( $\pm$ 0.86)	2.78 ( $\pm$ 1.12)
The Hague	9.06 ( $\pm$ 7.97)	10.45 ( $\pm$ 8.83)	5.93 ( $\pm$ 4.14)	7.88 ( $\pm$ 5.57)	1.98 ( $\pm$ 1.28)	3.01 ( $\pm$ 1.89)	1.98 ( $\pm$ 1.29)	3.01 ( $\pm$ 1.89)	2.06 ( $\pm$ 1.58)	2.99 ( $\pm$ 2.07)
Rotterdam	57.95 ( $\pm$ 40.14)	67.40 ( $\pm$ 47.14)	22.36 ( $\pm$ 18.14)	31.21 ( $\pm$ 24.98)	8.57 ( $\pm$ 6.31)	12.37 ( $\pm$ 8.77)	8.58 ( $\pm$ 6.35)	12.39 ( $\pm$ 8.82)	8.13 ( $\pm$ 4.77)	11.72 ( $\pm$ 6.59)

Table 6: Average city-level performance comparison for the 60 min. horizon (the lowest error in bold type).

District	Avg. Pop.	r‑MoDE‑Boost		Local XGBoost		TimeGPT
District	Avg. Pop.	MAE	RMSE	MAE	RMSE	MAE	RMSE
Bedrijvenpark Noord‑West	16	0.69	1.13	1.18	1.48	0.95	1.59
Charlois	248	8.64	13.29	7.76	11.56	12.52	17.11
Delfshaven	398	9.73	13.01	11.14	14.75	22.15	26.51
Feijenoord	385	8.43	11.23	8.67	11.30	14.53	18.01
Hillegersberg‑Schiebroek	435	14.08	21.50	12.03	18.22	17.06	23.41

Table 7: Performance comparison of the r-MoDE-Boost, local XGBoost models, and TimeGPT for the 60 min. forecasting horizon across Rotterdam districts (the lowest error in bold type).

Dataset	ASTN		MoDE-Boost
Dataset	MAE	RMSE	MAE	RMSE
CitiNYC	1.20	2.22	1.13	1.78
Divvy	0.77	1.43	0.76	1.25

Table 8: Performance comparison of r-MoDE-Boost and ASTN (as reported in Feng and Liu (2024) ) models on CitiNYC and Divvy datasets, measured by MAE and RMSE. Note: in this experiment, the error scores are reported in standardized units (demand counts).

Figure 7: Inference time of the MoDE-Boost model (both variants) over three different datasets and a 60 min. forecasting horizon, running at an Edge environment.

Spatiotemporal Forecasting

Spatiotemporal forecasting predicts how quantities change across locations and time, from traffic and weather to disease and environmental systems.

See all 423 papers

STAGformer: A Spatio-temporal Agent Graph Transformer for Micro Mobility Demand Forecasting
Jul 7, 2026Ye ZihaoAttention Mechanisms Spatiotemporal Forecasting
Department of System Engineering City University of Hong Kong
Learning to Transfer Across Modes: Towards Unified Urban Mobility Forecasting
Aug 28, 2026Yixuan Zhao, Man LuoRepresentation Learning Transfer Learning
University of Exeter Department of Computer Science Exeter, United Kingdom
TD-STGT: A Spatio-Temporal Graph Transformer for Mobile Traffic Demand Forecasting
Sep 6, 2026Mohamad Alkadamani, Halim YanikomerogluTraffic Forecasting Spatiotemporal Forecasting
Communications Research Centre Canada, Ottawa, ON, Canada · Department of Systems and Computer Engineering, Carleton University, Ottawa, ON, Canada

Jul 7, 2026J/K move · Enter open · S save

STAGformer: A Spatio-temporal Agent Graph Transformer for Micro Mobility Demand Forecasting

Ye Zihao

Department of System Engineering City University of Hong Kong

Attention Mechanisms Spatiotemporal Forecasting Time Series Transformers Spatiotemporal GNNs

Open paper AskPDF arXiv

Accurate station-level demand forecasting is essential for the efficient operation of bike-sharing systems, yet it remains challenging due to complex spatio-temporal dependencies and the large scale of urban networks. This paper presents STAGformer, a Spatio-Temporal Agent Graph Transformer that achieves efficient global modeling with linear computational complexity. The model introduces a two-step agent attention mechanism, where a small set of learnable spatial and temporal agent tokens first aggregate global information and then broadcast it back to individual stations and time steps, effectively capturing long-range interactions while reducing the quadratic cost of standard self-attention to O(NT). STAGformer integrates four core modules: a spatio-temporal encoder that fuses dynamic node features with external contextual factors (weather, time, points of interest), a graph propagation module for spatial neighbor aggregation, a temporal convolution module for local pattern extraction, and the agent attention module for global dependency modeling. Extensive experiments on two real-world datasets -- NYC Citi-Bike and Chicago Divvy-Bike -- demonstrate that STAGformer consistently outperforms state-of-the-art baselines across multiple prediction horizons, achieving significant improvements in both RMSE and MAE. Ablation studies validate the contribution of each component, with the agent attention mechanism proving critical for modeling global spatio-temporal dependencies.

Abstract

Spatiotemporal Forecasting

Explore similar work

STAGformer: A Spatio-temporal Agent Graph Transformer for Micro Mobility Demand Forecasting

Learning to Transfer Across Modes: Towards Unified Urban Mobility Forecasting

TD-STGT: A Spatio-Temporal Graph Transformer for Mobile Traffic Demand Forecasting