Modelo aumentado de árbol de decisión utilizando mapas autoorganizados

Abstract

A key aspect in the process of Knowledge Discovery in Databases (KDD), is the understandability and credibility of models generated by inductive learning schemes. This article explores the application of Self-Organizing Maps (SOM) technique on a model of decision tree, to achieve enhanced visualization of the model. Representations of visual perception model, together with data and patterns are established, based on a schema called VAM-MD, to support exploration and visual analysis efficiently during tuning stage for Data Mining Model. This seeks to answer generic questions about the inner workings of the model and to achieve better understanding. This proposal was implemented through a software prototype, where a set of visual elements may be applied to data from each node in the tree, appropriately selected to complement the visualization of the generated model. Additionally, the user has several mechanisms of interaction that enable for exploration of each component of the model. Finally, the results from a controlled experiment, conducted on two user groups who used the WEKA software and the experimental prototype, for a data mining task on a previously prepared data were analyzed. Preliminary analysis of the results obtained allow, on one hand empirically corroborate the utility of using the SOM technique to visually enhance a decision tree and on the other subjectively estimate their efficiency in supporting the understanding of the model generated.