Using Decision Trees for Knowledge-Assisted Topologically Structured Data Analysis

Supervised learning of an ensemble of randomized trees is considered to recognize classes of events in topologically structured data (e.g. images or time series). We are primarily interested in classification problems that are characterized by severe scarcity of the training samples. The main idea of our paper consists in favoring the selection of attributes that are known to efficiently discriminate the minority class in those nodes of the tree that are close to the leaves and where classes are represented by a small number of training examples. In practice, the knowledge about the ability of an attribute to discriminate the classes represented in a particular node is either provided by an expert or inferred based on a pre-analysis of the entire initial training set. The experimental validation of our approach considers sign language and human behavior recognition. It reveals that the proposed knowledge- assisted tree induction mechanism efficiently compensates for the shortage of the training samples, and significantly improves the tree classifier accuracy in such scenarios.