Two-Step Cluster Based Feature Discretization of Naive Bayes for Outlier Detection in Intrinsic Plagiarism Detection

Intrinsic plagiarism detection is the task of analyzing a document with respect to undeclared changes in writing style which treated as outliers. Naive Bayes is often used to outlier detection. However, Naive Bayes has assumption that the values of continuous feature are normally distributed where this condition is strongly violated that caused low classification performance. Discretization of continuous feature can improve the performance of Naive Bayes. In this study, feature discretization based on Two-Step Cluster for Naive Bayes has been proposed. The proposed method using tf-idf and query language model as feature creator and False Positive/False Negative (FP/FN) threshold which aims to improve the accuracy and evaluated using PAN PC 2009 dataset. The result indicated that the proposed method with discrete feature outperform the result from continuous feature for all evaluation, such as recall, precision, f-measure and accuracy. The using of FP/FN threshold affects the result as well since it can decrease FP and FN; thus, increase all evaluation.