Overfitting Reduction in Convex Regression

Résumé 0

Convex regression is a method for estimating an unknown function $f_0$ from a data set of $n$ noisy observations when $f_0$ is known to be convex. This method has played an important role in operations research, economics, machine learning, and many other areas. It has been empirically observed that the convex regression estimator produces inconsistent estimates of $f_0$ and extremely large subgradients near the boundary of the domain of $f_0$ as $n$ increases. In this paper, we provide theoretical evidence of this overfitting behaviour. We also prove that the penalised convex regression estimator, one of the variants of the convex regression estimator, exhibits overfitting behaviour. To eliminate this behaviour, we propose two new estimators by placing a bound on the subgradients of the estimated function. We further show that our proposed estimators do not exhibit the overfitting behaviour by proving that (a) they converge to $f_0$ and (b) their subgradients converge to the gradient of $f_0$, both uniformly over the domain of $f_0$ with probability one as $n \rightarrow \infty$. We apply the proposed methods to compute the cost frontier function for Finnish electricity distribution firms and confirm their superior performance in predictive power over some existing methods.