Redescription Mining with Multi-target Predictive Clustering Trees

摘要

Redescription mining is a field of knowledge discovery that aims to find different descriptions of subsets of elements in the data by using two or more disjoint sets of descriptive attributes. The ability to find connections between different sets of descriptive attributes and provide a more comprehensive set of rules makes it very useful in practice. In this work, we introduce redescription mining algorithm for generating and iteratively improving a redescription set of user defined size based on multi-target Predictive Clustering Trees. This approach uses information about element membership in different generated rules to search for new redescriptions and is able to produce highly accurate, statistically significant redescriptions described by Boolean, nominal or numeric attributes. As opposed to current tree-based approaches that use multi-class or binary classification, we explore benefits of using multi target classification and regression to create redescriptions. The process of iterative redescription set improvement is illustrated on the dataset describing 199 world countries and their trading patterns. The performance of the algorithm is compared against the state of the art redescription mining algorithms.