Visual Analytics - Decision Trees (2)

The second part of the decision tree lecture elaborates on the visualization of decision trees and the interactive exploration that is essential if the trees exceed a certain size. This discussion is based on frequently used evaluation criteria. The basic technique of node-link diagrams can be refined by revealing how many datasets are represented in the different branches of the tree. Also, histogram-based representations convey information related to the subsets of datasets that are shown in different parts of the tree. Layout computation, labelling, interactive pruning/expanding are aspects of the visual exploration. The Baobab view inspired by African trees is a landmark approach for the visual exploration of decision trees. We go on with thoughts on validation, e.g., how to analyze instances of misclassifications. Which patterns arise in a misclassification? A recent paper introduced the use of Pareto-optimization to find decision trees that are optimal in terms of complexity and accuracy. The explicit visualization of the Pareto front is an important visual aid. Finally, regression trees are introduced. They generate a scalar value as output in contrast to the discrete, often binary classification performed by a decision tree. Chapters: 00:00 - Advanced Visualization 21:26 - Interaction 26:42 - Baobab 34:43 - Automatic vs. Semi-interactive Model Construction 39:36 - Validation 50:21 - Related Concepts 54:13 - Summary and References