Abstract
Quantile regression is a powerful tool for non-parametric uncertainty estimation in univariate settings. However, extending it to multivariate targets is challenging due to the lack of a natural ordering in higher dimensions. This paper introduces Quantile Surfaces (QS), a method that bypasses this limitation by transforming the problem into a radial coordinate system. By predicting distances from a central point across various directions, the proposed Quantile Surface Regression Neural Network (QSNN) can capture complex, non-Gaussian uncertainty shapes. The method is validated on synthetic data, wind power generation, and autonomous vehicle path planning, showing significant improvements over traditional Gaussian baselines.
Introduction: The Challenge of Multivariate Uncertainty
In many critical forecasting tasks—such as predicting the trajectory of a cyclist or the power output of a network of wind farms—a single point prediction is insufficient. Decision-makers require a clear understanding of uncertainty to manage risks effectively. While Quantile Regression (QR) is the standard for non-parametric uncertainty in 1D, it fails in multivariate spaces because there is no objective way to define "greater than" or "less than" for coordinate pairs.
Existing multivariate methods often rely on Gaussian assumptions, which struggle with the skewed or irregular distributions common in real-world data. Quantile Surfaces address this by defining quantiles based on the distance from a central prediction, allowing for flexible, "star-shaped" probability envelopes.
Methodology: From Points to Surfaces
The Quantile Surface approach employs a two-stage estimation process:
- Central Tendency Estimation: A deterministic model (e.g., a standard neural network) first predicts the "center" of the distribution, such as the mean or median.
- Quantile Surface Estimation: Instead of predicting raw coordinates, the model is trained to predict the distance ($d$) from the center for any given direction ($\mathbf{u}$).
By shifting to a radial representation, the model recovers a 1D ordering: for any specific angle, quantiles can be defined based on distance.
The QSNN and Directional CRPS
The authors implement this via the Quantile Surface Regression Neural Network (QSNN). This architecture takes input features and a directional vector to output predicted distances for multiple quantile levels (e.g., 10%, 50%, 90%). To evaluate these forecasts, the paper introduces Directional CRPS, a metric that assesses the model's skill along the radial directions where actual observations occur, providing a computationally efficient alternative to standard multivariate scoring.
Experimental Validation
The researchers tested the QS framework across three domains:
- Synthetic Data: Proving the model's ability to capture Gaussian, skewed, and conditional distributions.
- Renewable Energy: Using the EuroPianWindFarm dataset to model spatial dependencies between 54 European wind farms.
- Vulnerable Road Users: Analyzing 1,056 cyclist trajectories at an urban intersection to provide reliable safety envelopes for autonomous systems.
Key Results and Insights
The "Aha" moment of this research is the use of radial ordering. By measuring uncertainty outward from a central point, "outward" becomes a universal proxy for extremity. This allows the QSNN to wrap itself around highly skewed data, such as "teardrop" shaped distributions, which Gaussian models fail to represent.
- Wind Power: The QSNN outperformed Gaussian baselines by over 80% in skill (Directional CRPS), effectively capturing spatial correlations across distances ranging from 20km to 1400km.
- Cyclist Trajectories: For short-term horizons (0.2s), the QSNN showed a 99% improvement in skill over traditional methods, providing much more reliable "safety envelopes."
- Reliability: Q-Q plots confirmed that the predicted 90% surfaces contained the true observations approximately 90% of the time, even in heavily skewed scenarios.
Limitations and Future Directions
While powerful, the current QS approach assumes a star-shaped distribution, meaning it is primarily suited for unimodal (single-peak) outcomes. It cannot currently handle multimodal scenarios, such as a cyclist choosing between two distinct paths at a fork. Future work aims to integrate Bayesian Neural Networks to capture epistemic uncertainty and utilize trigonometric basis functions to ensure perfectly smooth surfaces across all angles.
Summary updated