An Interpretable Classification Model Based on Characteristic Element Extraction

The process of a classification application is usually dynamic and long. During the process of an application, better classification application effect can be acquired by enlarging and adjusting the training dataset continuously, for example, modifying the wrong labels of original instances. For this kind of dynamic classification applications, how to build an interpretable classifier which can help domain experts to understand each label's meanings reflected from the dataset, then to compare and discriminate them with their own mastered domain knowledge, and finally to adjust and optimize the training set to enhance the effect of classification applications, is a neglected but worth studying issue. Therefore, an interpretable classification model based on characteristic element extraction is proposed in this paper. The proposed classifier is constructed by extracting positive and negative characteristic elements for all class labels which can intuitively reflect their instinct characteristics. Thus, it has high interpretability obviously and can effectively help domain experts optimize classification effect. At the same time, experiment results show that our classifier also has higher accuracy compared with other kinds of classical classifiers. Consequently, the classification model proposed in this paper is effective and efficient, especially in practical applications.