Explainable classification by learning human-readable sentences in feature subsets

We propose a new methodology (Sentences in Feature Subsets, i.e., SiFS) to mine human-readable decision rules from empirical data sets. Unlike opaque classifiers obtained using deep learning, the proposed methodology derives decision rules that are compact and comprised of Boolean logic sentences involving subsets of features in the input data. For this purpose, we develop a new classifier model defined in terms of sets of inequalities among selected features in the input data. To empirically derive suitable inequalities from training data, our approach combines a differentiable representation of sets of Boolean logic sentences, gradient-based optimization of coefficients in the inequalities, a genetic-based algorithm for selection of the subsets of features, and a “goodness” model of sentences to prune and down-select sentences. We present results on synthetic and real-world benchmark datasets to demonstrate efficacy of SiFS in deriving human-readable decision rules. It is seen that SiFS achieves comparable accuracies to the best among various other classification algorithms (accuracies of 95% to 100% on several datasets, F₁ scores between 0.95 and 1.0), reasonable computation times (training times of a few seconds for considered datasets), and compact human-readable decision rules (between 1 to 10 sentences of 3 words or less for considered datasets).