CoCoNet: : Convex Conformal Network for Multi-Dimensional Uncertainty Regions

In the world of data science, predicting a single number is often straightforward, but predicting a cluster of related outcomes is a different challenge entirely. When a model forecasts the future, it rarely offers a single, perfect point; instead, it must provide a range of plausible outcomes to account for uncertainty. For simple predictions, this range might be a single interval. However, when predicting multiple things at once—such as the future position of a car in three dimensions, or the temperature, humidity, and wind speed at a weather station—the uncertainty forms a complex, multi-dimensional shape. The goal is to draw a boundary around these possibilities that is tight enough to be useful for decision-making but wide enough to be trustworthy. If the boundary is too loose, it offers no practical guidance; if it is too tight, it risks missing the actual outcome. The challenge lies in shaping this boundary so that it matches the true, often irregular, geometry of the uncertainty, rather than forcing it into a rigid, pre-defined shape like a box or a sphere.

A team of researchers at the Georgia Institute of Technology has developed a new method called CoCoNet to solve this problem. Their work focuses on creating prediction regions that are not only mathematically guaranteed to be correct a certain percentage of the time but are also as small and efficient as possible. Traditional methods often rely on fixed geometric templates, such as stretching an ellipse or a rectangle to fit the data. While these shapes are easy to work with, they fail when the true uncertainty is skewed, lopsided, or has a complex, irregular form. The researchers found that by teaching a computer to learn the shape of the uncertainty directly from the data, while strictly enforcing a rule that the shape must be convex—meaning it has no inward dents or holes—they could create prediction zones that adapt perfectly to the data's structure.

The core of their approach involves a two-step process. First, the system learns a scoring function that measures how "surprising" a potential outcome is compared to a standard prediction. Crucially, this scoring function is designed to be convex, ensuring that the resulting region of plausible outcomes is a single, solid shape without gaps. The researchers tested two versions of this system: one that builds shapes using a series of connected flat planes, similar to a folded piece of paper, and another that uses a highly flexible neural network architecture capable of molding the shape in almost any way. They then calibrated these learned shapes using a separate set of data to ensure they met the required coverage guarantee. This means that if the system is set to be 90% confident, the actual outcome will fall inside the predicted region 90% of the time, regardless of the underlying data distribution.

To test their method, the researchers applied it to thirty different real-world problems, ranging from predicting the movement of autonomous vehicles to estimating energy consumption in buildings and analyzing chemical properties of molecules. These tasks involved predicting anywhere from two to sixteen different outcomes simultaneously. They compared their new method against several existing techniques, including those that produce simple rectangles, ellipses, or other standard shapes. The results were striking. The most flexible version of their system, which uses the neural network to learn the shape, produced the most efficient prediction regions in twenty-two out of the thirty tasks. In many cases, the regions it generated were significantly smaller than those produced by the best existing methods, meaning it could provide much tighter, more useful bounds on uncertainty without sacrificing reliability.

The study also explored what happens when the underlying uncertainty does not fit a neat pattern. In synthetic tests where the true uncertainty took the form of a triangle or a trapezoid, standard methods struggled, often producing large, inefficient regions that wasted space to cover the correct area. In contrast, the new method learned the exact geometry of these shapes, creating prediction regions that hugged the data closely. Even in cases where the uncertainty was roughly elliptical, the new method performed as well as or better than methods specifically designed for ellipses, proving its versatility. The researchers demonstrated that by allowing the shape to be learned from the data rather than imposed from the outside, the system could capture subtle asymmetries and dependencies between the different outcomes that other methods missed.

This work matters because the quality of a prediction region directly impacts the decisions made based on it. In fields like robotics, finance, or climate modeling, knowing the precise limits of uncertainty allows for safer and more effective planning. A prediction region that is unnecessarily large forces decision-makers to prepare for worst-case scenarios that are unlikely to happen, leading to inefficient resource use. Conversely, a region that is too small might miss the actual outcome, leading to catastrophic failures. By providing a way to generate prediction regions that are both mathematically sound and geometrically efficient, this new method offers a more reliable tool for navigating uncertainty in complex, multi-dimensional environments. The researchers have made their code and data available, allowing others to apply this approach to their own problems, bridging the gap between theoretical guarantees and practical, high-performance uncertainty quantification.