Rule-Based Classifier Using Bernoulli Mixture Model for Utilizing Published Rules

Abstract

Decision rules for classification problem are requested to be open for public inspection to ensure the fairness. We may construct an effective and useful classifier by combining the rules published by different organizations with rules induced from a dataset at hand. However, the published rules may include some special characters caused by territorial character, cultural ethnic and religious, different age strata bias and so on. Under this consideration, we should verify how to utilize the published rules appropriately. Few studies have devoted to the combination of the published rules with rules induced from a dataset at hand. In this paper, we propose a mixture model approach to constructing an effective binary rule-based classifier by utilizing the published rules. We apply the idea of the mixture model in order to utilize both the published rules and the induced rules from dataset at hand. By applying LERS algorithm to each set of rules, we obtain two normalized score distributions over decision classes. EM algorithm is then used for estimating the appropriate mixture ratios for constructing a classifier. Numerical experiments are conducted to examine the performance of the proposed method. The proposed method is compared with four alternative methods in four real datasets. The effectiveness of the proposed method is demonstrated by the numerical experiments.