Simple Mimetic Classifiers

The combination of classifiers is a powerful tool to improve the accuracy of classifiers, by using the prediction of multiple models and combining them. Many practical and useful combination techniques work by using the output of several classifiers as the input of a second layer classifier. The problem of this and other multi-classifier approaches is that huge amounts of memory are required to store a set of multiple classifiers and, more importantly, the comprehensibility of a single classifier is lost and no knowledge or insight can be acquired from the model. In order to overcome these limitations, in this work we analyse the idea of “mimicking” the semantics of an ensemble of classifiers. More precisely, we use the combination of classifiers for labelling an invented random dataset, and then, we use this artificially labelled dataset to re-train one single model. This model has the following advantages: it is almost similar to the highly accurate combined model, as a single solution it requires much fewer memory resources, no additional validation test must be reserved to do this procedure and, more importantly, the resulting model is expressed as a single classifier in terms of the original attributes and, hence, it can be comprehensible. First, we illustrate this methodology using a popular data-mining package, showing that it can spread into common practice, and then we use our system SMILES, which automates the process and takes advantage of its ensemble method.