Learning metrics and discriminative clustering

Abstract

In this work methods have been developed to extract relevant information from large, multivariate data sets in a flexible, nonlinear way. The techniques are applicable especially at the initial, explorative phase of data analysis, in cases where an explicit indicator of relevance is available as part of the data set.

The unsupervised learning methods, popular in data exploration, often rely on a distance measure defined for data items. Selection of the distance measure, part of which is feature selection, is therefore fundamentally important.

The learning metrics principle is introduced to complement manual feature selection by enabling automatic modification of a distance measure on the basis of available relevance information. Two applications of the principle are developed. The first emphasizes relevant aspects of the data by directly modifying distances between data items, and is usable, for example, in information visualization with the self-organizing maps. The other method, discriminative clustering, finds clusters that are internally homogeneous with respect to the interesting variation of the data. The techniques have been applied to text document analysis, gene expression clustering, and charting the bankruptcy sensitivity of companies.

In the first, more straightforward approach, a new local metric of the data space measures changes in the conditional distribution of the relevance-indicating data by the Fisher information matrix, a local approximation of the Kullback-Leibler distance. Discriminative clustering, on the other hand, directly minimizes a Kullback-Leibler based distortion measure within the clusters, or equivalently maximizes the mutual information between the clusters and the relevance indicator. A finite-data algorithm for discriminative clustering is also presented. It maximizes a partially marginalized posterior probability of the model and is asymptotically equivalent to maximizing mutual information.