Evzenie Coupkova

Evzenie Coupkova, Mireille Boutin

Given a classification problem and a family of classifiers, the Rashomon ratio measures the proportion of classifiers that yield less than a given loss. Previous work has explored the advantage of a large Rashomon ratio in the case of a finite family of classifiers. Here we consider the more general case of an infinite family. We show that a large Rashomon ratio guarantees that choosing the classifier with the best empirical accuracy among a random subset of the family, which is likely to improve generalizability, will not increase the empirical loss too much. We quantify the Rashomon ratio in two examples involving infinite classifier families in order to illustrate situations in which it is large. In the first example, we estimate the Rashomon ratio of the classification of normally distributed classes using an affine classifier. In the second, we obtain a lower bound for the Rashomon ratio of a classification problem with a modified Gram matrix when the classifier family consists of two-layer ReLU neural networks. In general, we show that the Rashomon ratio can be estimated using a training dataset along with random samples from the classifier family and we provide guarantees that such an estimation is close to the true value of the Rashomon ratio.

Mireille Boutin, Evzenie Coupkova

The generalization gap of a classifier is related to the complexity of the set of functions among which the classifier is chosen. We study a family of low-complexity classifiers consisting of thresholding a random one-dimensional feature. The feature is obtained by projecting the data on a random line after embedding it into a higher-dimensional space parametrized by monomials of order up to k. More specifically, the extended data is projected n-times and the best classifier among those n, based on its performance on training data, is chosen. We show that this type of classifier is extremely flexible as, given full knowledge of the class conditional densities, under mild conditions, the error of these classifiers would converge to the optimal (Bayes) error as k and n go to infinity. We also bound the generalization gap of the random classifiers. In general, these bounds are better than those for any classifier with VC dimension greater than O(ln n). In particular, the bounds imply that, unless the number of projections n is extremely large, the generalization gap of the random projection approach is significantly smaller than that of a linear classifier in the extended space. Thus, for certain classification problems (e.g., those with a large Rashomon ratio), there is a potntially large gain in generalization properties by selecting parameters at random, rather than selecting the best one amongst the class.

Mireille Boutin, Evzenie Coupkova

Clustering aims to divide a set of points into groups. The current paradigm assumes that the grouping is well-defined (unique) given the probability model from which the data is drawn. Yet, recent experiments have uncovered several high-dimensional datasets that form different binary groupings after projecting the data to randomly chosen one-dimensional subspaces. This paper describes a probability model for the data that could explain this phenomenon. It is a simple model to serve as a proof of concept for understanding the geometry of high-dimensional data. We start by building a rescaled multivariate Bernouilli model (stretched hypercube) so to create several overlapping grouping structures in the data. The size of each scaling parameter is related to the likelihood of uncovering the corresponding grouping by random 1D projection. Clusters in the original space are then created by adding noise to this cluster-free model. In high dimension, these clusters would hardly be observable given a sample set from the distribution because of the curse of dimensionality, but the binary groupings are clear. Our construction makes it clear that one needs to make a distinction between "groupings" and "clusters" in the original space. It also highlights the need to interpret any clustering found in projected data as merely one among potentially many other groupings in a dataset.