Rule Extraction from Opaque Models-- A Slightly Different Perspective

Abstract

When performing predictive modeling, the key criterion is always accuracy. With this in mind, complex techniques like neural networks or ensembles are normally used, resulting in opaque models impossible to interpret. When models need to be comprehensible, accuracy is often sacrificed by using simpler techniques directly producing transparent models; a tradeoff termed the accuracy vs. comprehensibility tradeoff. In order to reduce this tradeoff, the opaque model can be transformed into another, interpretable, model; an activity termed rule extraction. In this paper, it is argued that rule extraction algorithms should gain from using oracle data; i.e. test set instances, together with corresponding predictions from the opaque model. The experiments, using 17 publicly available data sets, clearly show that rules extracted using only oracle data were significantly more accurate than both rules extracted by the same algorithm, using training data, and standard decision tree algorithms. In addition, the same rules were also significantly more compact; thus providing better comprehensibility. The overall implication is that rules extracted in this fashion will explain the predictions made on novel data better than rules extracted in the standard way; i.e. using training data only.