Automatic Context Selection in Explainable Support Vector Machine Classification

Abstract

The contextualization of knowledge models has recently been proposed for improving the interpretability of predictions made by explainable support vector machine classifiers. An advantage attributed to the improved classifiers is the capability of selecting the context of the resulting predictions. However, this advantage is inapplicable in situations where a data set with unknown context is provided for evaluation. To make the advantage applicable in such situations, a method for automatic context selection is proposed in this paper. The method uses a reduced number of labeled samples contained in the data set with unknown context to select, from an arrangement of contextualized models, the model whose context is the most similar to the unknown context. The selected context is then used for making the predictions of the remaining elements in the provided data set. Illustrative examples show how the proposed method helps to improve the interpretability and accuracy of the resulting predictions.