Impacts of Mutual Information and Discretization Methods in Associative Classifications

Objective: Data Pre-processing plays a vital role in associative classification to influence the learning outcome. Feature selection and discretization methods contribute more in implementing associative classification especially on gene expression data. The discretization method determines discrete intervals and these intervals are impacted on number of itemsets to create a transaction format for generating class association rules. This paper aims to find a suitable combination of feature selectionand discretization methods to generate association rules for classifying the cancer gene expression data. Methods: Three mutual information based features selection methods joint impurity filter (JIM), joint mutual information filter (JMI) and third-order joint mutual information filter(JMI3) are considered and five supervised discretization methods minimum description length principles (MDLP), class-attribute inter dependence maximization (CAIM), class-attribute contingency coefficient (CACC), an autonomous discretization algorithm (AMEVA) and chi-squared Merge (Chi-Merge) are considered for selecting the best combination. This combination is used in CMAR and CBA associative classification. Findings: CMAR gives 100% accuracy for the cancer gene expression datasets GSE15781 and GSE2685 for any combination of the assumed filter method and supervised discretization method. The considered evaluation metrics brier score, F-beta score, F1 score and AUC curve are also shown that CMAR provides better results when the information based filter method is combined with the supervised discretization. Novelty: Proposed method is used to find the informative genes using the opt mutual information method and discretization method to predict the gene is affected by cancer or not. The proposed method diagnoses diseases from microarray gene expression data using associative classification techniques for classifying the cancer gene expression data. Keywords:Gene expression, Mutual information, Discretization, Associative classification, CBA, CMAR