Categorization by simplicity: a minimum description length approach to unsupervised clustering
Emmanuel M. Pothos, Nick Chater · 2001
Abstract There is a strong intuition that one important factor in determining psychological categories is that they should group similar items together — that categories should be seen as running along default lines in a psychological similarity space. To make this idea precise requires finding some ‘objective’ criterion for determining a ‘good’ classification, given a set of similarity data. This chapter provides such a criterion, based on an application of a simplicity principle, that can be viewed as a general criterion for cognition. To illustrate the approach, it addresses the specific illustrative problem of dividing a set of items into groups, on the basis of data consisting of pairwise similarities between the items. A simplicity principle is used to assess the relative goodness of different clusterings on the same data set: a particular classification is good to the extent that it provides a short encoding of the similarity information.