Untangling webs of interactions in classification models Popular Science Summary

Rule-based classifiers have one major advantage over other classes of supervised learning algorithms: interpretability. They provide a means to read into a model and find how the features co-act in order to come to a classification outcome. This in turn enables the researcher to visualize the feature interactions and evaluate the key features that discern between different decision classes. The rules generated from these algorithms, however, can be very many and their analysis is not trivial. This is where proper visualization techniques enable the researcher to filter out clutter and see only important relationships. In addition, the next natural step for genomic data is to find out relationships between the interacting genes and biological networks is always a good starting place. In this study, we introduce VisuNet, a highly interactive, web-based tool for visualization of feature interactions in rule-based classifiers as well as annotation of genomic data with information on biological networks involved. VisuNet can be used with any rule-based classifiers such as decision trees and Rough-Sets, or any model from which rules can be extracted. The tool is hosted online at http://bioinf.icm.uu.se/~visunet/.