L1-Regulated Feature Selection and Classification of Microarray Cancer Data Using Deep Learning

Abstract

The microarray cancer data obtained through microarray technology poses a lot of challenges during classification since the sample size is very small and the dimensionality of the data is very high. It is noticed that usually, the number of classes in multiclass datasets are highly imbalanced. In order to reduce the dimensionality thereby enabling accurate classification, in this work, we propose an L1-regulated feature selection and deep learning is applied for classification. The L1-regulated feature selection is based on Linear Support Vector Machine (LSVM) which is characterized by adding a penalty term to the prediction error in order to reduce the weight of the irrelevant features and to make the relevant features having nonzero weights. For classification purpose, deep learning neural network is initialized with sigmoid activation function in the input and hidden layers and to accommodate multiclass classification, the softmax activation function is used in the output layer. In order to demonstrate the suitability of the proposed approach, experiments are conducted on the six numbers of standard multiclass cancer datasets and to argue the predictive capability of the proposed approach, experiments are conducted on imbalanced class datasets such as 5-class lung cancer dataset, and 4-class Leukemia cancer dataset. Comparative study is also provided with state-of-the-art approaches and the results are presented considering classification accuracy, precision, recall, f-measure, confusion matrix, average precision, and ROC metrics to exhibit the performance of the proposed approach.