📊 statistics

Robust OT-Guided Generative Residual Domain Adaptation for Bike-Sharing Demand Prediction under Temporal Domain Shift

This paper proposes Gen-ROTDA, a robust optimal transport-guided generative residual domain adaptation framework that effectively predicts bike-sharing demand under temporal domain shifts by transferring residual patterns and trimming high-cost transport matches, outperforming various baselines in noisy and multi-year scenarios.

Original authors: Yiming Ma

Published 2026-05-25

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Yiming Ma

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a city planner trying to guess how many people will rent bikes from a specific station at a specific hour. You have a very smart computer model trained on data from last year (2025). You want to use this model to predict demand for next year (2026).

The problem is that people change. Maybe a new office building opened, the weather patterns shifted, or commuters changed their habits. If you just use last year's model directly, it will make mistakes because the "rules" of the city have changed. This is called Temporal Domain Shift.

This paper introduces a new method called Gen-ROTDA to fix this. Here is how it works, explained through simple analogies:

1. The "Anchor" and the "Drift" (Decomposition)

Imagine the bike demand is like a boat on a river.

The Anchor: Some things about the boat stay the same no matter the year. The station's location, the time of day, and whether it's a weekend are like the boat's anchor. These are stable. The authors build a simple model just for these stable parts.
The Drift (Residual): The part that changes is the "drift"—the extra demand caused by new events, weird weather, or shifting habits. This is the messy, unpredictable part.

Instead of trying to predict the whole boat movement again, the authors' method only tries to predict the drift (the residual). They subtract the stable "anchor" prediction from the real data and only try to fix the messy leftover part. This makes the problem much easier.

2. The "Translator" (The Generator)

Now, imagine you have a dictionary from 2025 (Source) and you need to speak 2026 (Target). The words are slightly different.

The paper uses a Generator (a small AI translator). It takes the 2025 "drift" data and tweaks it just enough so it looks more like 2026 data.
Crucially, it doesn't rewrite the whole story; it just adds a little "translation layer" so the 2025 data fits better with 2026 patterns without losing the original meaning.

3. The "Smart Matchmaker" with a "Safety Net" (Robust Optimal Transport)

This is the most important part. To teach the model, you need to match 2025 data points with 2026 data points.

Standard Matching (OT): Imagine a matchmaker trying to pair up 2025 and 2026 data points based on how similar they are.
The Problem: Sometimes, the 2026 data has "noise"—weird records caused by a sudden storm, a data glitch, or a one-time event. A standard matchmaker might try to force a match between a normal 2025 day and a weird 2026 day, which ruins the prediction.
The Solution (Robust OT): The authors added a Safety Net. Before finalizing the matches, the system looks at the "cost" of each match. If a match is too weird or expensive (like trying to pair a sunny day with a hurricane day), the system cuts that match and ignores it. It only keeps the "good" matches to train the model.

What Did They Find?

The authors tested this on Citi Bike data in New York City from 2021 to 2026.

Best at the Main Task: For predicting 2026 based on 2025, their method (Gen-ROTDA) made the fewest errors compared to other methods.
The "Safety Net" is Key: When they intentionally added "garbage" or weird data to the 2026 set (simulating bad weather or data errors), the standard methods got confused and failed. Gen-ROTDA, thanks to its "Safety Net" (Robust OT), stayed calm and accurate.
Not a Magic Bullet: While it was the best at handling the "noisy" data and was the top method among the "Optimal Transport" family, simple methods like just "fine-tuning" (re-training the model slightly on new data) were still very strong competitors overall.

The Bottom Line

The paper argues that when predicting bike demand across different years, you shouldn't try to predict everything at once. Instead:

Separate the stable stuff from the changing stuff.
Translate the changing stuff to look like the new year.
Ignore the weird, noisy matches that don't make sense.

By doing this, the model becomes much more reliable, especially when the real-world data is messy or contains unexpected surprises.

Technical Summary: Robust OT-Guided Generative Residual Domain Adaptation for Bike-Sharing Demand Prediction

Problem Statement
Bike-sharing demand prediction models trained on historical station-hour data often degrade when deployed in subsequent years due to temporal domain shifts. Factors such as changing station popularity, commuting behaviors, rider composition, and operational patterns alter the underlying data distribution ( $P_s(X, Y) \neq P_t(X, Y)$ ). This challenge is exacerbated by the sparsity, skewness, and noise inherent in demand data, where a few stations and peak hours dominate prediction errors. Furthermore, target domains may contain abnormal records caused by unusual events or data quality issues, which can destabilize standard domain adaptation methods. The paper addresses the specific task of predicting March Citi Bike demand from 2021 to 2026, focusing on the 2025-to-2026 transfer as the primary evaluation scenario.

Methodology: Gen-ROTDA
The paper proposes Gen-ROTDA (Generative Robust Optimal Transport Domain Adaptation), a framework designed to handle temporal shifts through a combination of residual decomposition, feature generation, and robust transport alignment. The methodology consists of four key stages:

Anchor-Residual Decomposition:
Instead of transferring raw demand, the method decomposes the log-transformed demand ( $z = \log(1+Y)$ ) into two components:
- Anchor Component: A target-domain model $a_\phi(A)$ trained on labeled target samples using stable spatial and calendar features (e.g., coordinates, hour, day of week). This captures the stable station-time structure.
- Residual Component: The difference $r = z - a_\phi(A)$ , which contains the year-specific dynamics. Domain adaptation is applied exclusively to this residual component.
Label-Preserving Residual Feature Generator:
A deterministic generator $G_\theta(T) = T + h_\theta(T)$ (where $h_\theta$ is a small MLP) transforms source residual features to resemble target features. Unlike stochastic GANs, this generator aims to keep transformed features close to the original representation while aligning distributions. The training objective combines:
- Distribution alignment loss (MMD-style) between generated source and target features.
- Identity regularization to prevent excessive feature movement.
- Label preservation and supervision losses to ensure demand-relevant information is retained.
Robust Optimal Transport (OT) Alignment:
After feature generation, the method aligns source and target distributions using Optimal Transport. To address the sensitivity of standard OT to outliers, Gen-ROTDA employs a robust trimming step:
- Transport coupling entries are sorted by cost.
- High-cost matches (outliers) are removed until the retained mass reaches a specified threshold (e.g., 0.8).
- The remaining coupling is renormalized to compute barycentric projections. If a source row loses all mass, the original generated feature is retained as a fallback.
Final Prediction:
A residual predictor is trained on the transported source residuals and labeled target residuals (with target residuals up-weighted). The final demand prediction is reconstructed by adding the predicted residual to the anchor prediction and applying the inverse log transformation.

Key Contributions

Problem Formulation: The paper formulates cross-year bike-sharing prediction as a residual temporal domain adaptation problem, separating stable structural effects from transferable residual dynamics.
Framework Proposal: It introduces Gen-ROTDA, which uniquely combines a deterministic, label-preserving residual feature generator with robust OT trimming.
Comprehensive Evaluation: The method is evaluated on Citi Bike data (2021–2026) across main tasks, multi-year transfers, robustness tests against abnormal records, and ablation studies.

Experimental Results

Main Task (2025 $\to$ 2026): Gen-ROTDA achieved the lowest Mean Absolute Error (MAE) of 0.8625, outperforming the source-only baseline (0.8834) and the Sinkhorn OTDA variant (0.8633). While Gen-OTDA (without robust trimming) achieved the lowest RMSE and highest $R^2$ , Gen-ROTDA was superior in minimizing average absolute error, which is critical for sparse, skewed demand data.
Multi-Year Performance: Across adjacent-year and two-year transfer tasks, Gen-ROTDA was the best-performing method within the OT-family, slightly outperforming Sinkhorn OTDA, ROTDA, and Gen-OTDA. However, it did not surpass strong baselines like Fine-tuning and MMD adaptation in overall average MAE.
Robustness: Under injected abnormal target records (up to 20% contamination), Gen-ROTDA demonstrated superior stability. Its MAE increased only marginally (from 0.8625 to 0.8629), whereas non-robust variants (Sinkhorn OTDA and Gen-OTDA) showed significant degradation.
Ablation: The study confirmed that robust OT trimming is the primary driver of MAE improvement. The generator provided a modest additional benefit but was not the dominant factor in isolation.

Significance and Claims
The paper claims that robust transport is the critical component for handling noisy temporal transfers in urban mobility forecasting. By trimming high-cost transport matches, the method effectively mitigates the impact of abnormal target records that typically destabilize standard OT approaches.

The authors position Gen-ROTDA not as a universally superior predictor across all baselines (acknowledging that simple fine-tuning and MMD adaptation remain strong), but as a robust OT-family method that offers significant stability in the presence of data anomalies. The anchor-residual design is highlighted as a practical strategy to isolate stable structures, making the residual adaptation problem lower-dimensional and more tractable. The findings suggest that while generative feature transformation offers flexibility, its value is complementary and task-dependent, whereas robust alignment is essential for reliable temporal domain adaptation in noisy environments.