Classification of Multi-class Microarray Cancer Data Using Ensemble Learning Method

Nowadays, microarray cancer analysis is one of the top research areas in the field of machine learning, computational biology, and pattern recognition. Classifying cancer data into their respective class and its analysis plays a key role in diagnosis, identifying negative and positive cases as well as treatment in the case of binary classes. In the case of multi-class classification, the aim is to identify the type of cancer. The main challenge in microarray cancer datasets is the curse of dimensionality and lack of sufficient sample data. To overcome this problem, feature selection and dimensionality reduction are explored in identifying relevant features. In this work, we propose an ensemble learning method for multi-class cancer data classification. The Information Gain (IG) is used for feature selection which works by ranking attributes according to their relevance with respect to the class label. Three classifiers are used, namely k-Nearest Neighbor, Logistic Regression, and Random Forest. tenfold cross validation is applied to train and test the model. Experiments are conducted on the standard multi-class cancer datasets, namely Leukemia 3 class, Leukemia 4 class, Harvard Lung cancer 5 class, and MLL 3 class. To evaluate the performance of the model, various performance measures such as Classification Accuracy, F1-measure, and Area Under the Curve (AUC) are used. Confusion matrix is used to show whether or not samples are correctly classified. Comparison of each classifier's performance is presented on the basis of performance evaluation criteria. Significant performance improvement is observed in the results due to feature selection for three of the classifiers with the exception of random forest's performance on MLL Leukemia whose result is found to be good on the original dataset compared to the selected features. For the rest of the datasets, all classifiers registered better result due to feature selection.