Two-step verifications for multi-instance features selection: A machine learning approach

Abstract

Multi-instance features measurement is an important step in identifying characteristics that are bound to various experimental events. In biological data processing, a set of critical factors is responsible for several diseases. Computational simulation will help to design an optimal tool for cost-effective drug design. In this regard, the processing of big data is valuable for efficient simulation. Recent experimental results generate huge amounts of related data. In the current work, noisy data have been treated with three filtering techniques: cross-validated committees filtering (CVCF), iterative partitioning filtering (IPF) and ensemble filtering (EF). A comparison was made of these three filtering approaches. The filtered datasets were normalized. The repeated application of three normalization techniques removed the irregularities and structured the datasets. Wide ranges of comparisons were made among these three normalization techniques. After being appropriately structured, these normalized datasets were transformed accordingly with three different transformation processes: rank transformation, nominal to binary transformation and Box-Cox transformation. To prevent false positive and false negative outcomes of the experiments, certain key aspects were considered: accuracy, sensitivity and F-measures. Accuracy of the experiments relates to the level of precise detection of certain factors; specificity allows the selection of the dominant factors; and sensitivity and F-measures are the ratio between the training and testing datasets. Detailed experimental analysis included a comparison study of the four classifiers for the deoxyribonucleic acid (DNA) dataset.