Parameter Tuning onto Recurrent Neural Network and Long Short-Term Memory (RNN-LSTM) Network for Feature Selection in Classification of High-Dimensional Bioinformatics Datasets

Abstract

Feature selection helps with the selection of relevant features that are present in large number of features and ignores the remaining features that have little value on output feature set. Deep learning methods have been applied to select relevant features in the classification problem; however, the current approach (i.e., search strategies) to the learning of a parameter can either grow out of bound or shrink (they decay exponentially in the number of layers) at each time step (iteration) with the subsequent effect of inaccurate classification of features. To address this challenge of the current search strategies, we proposes an approach to the learning of a parameter for the classification problem based on the behavior of birds (i.e., kestrel bird). The proposed approach, bio-inspired approach, is modeled as a search algorithm which is then integrated with deep learning method. The integration enables learning of optimum parameter for feature selection in a classification problem. A benchmark dataset (i.e., bioinformatics dataset with continuous data attributes) from the Arizona State University was chosen because of its high dimensionality and its continuous data attribute nature. This dataset was used to test the proposed algorithm. The algorithm proposed was evaluated against comparative bio-inspired algorithms namely PSO, ACO, WSA-MP and BAT. The findings indicate that KSA produces minimum learning rate in five datasets out of nine datasets. While on the classification accuracy, KSA produces the highest accuracy of classification in four out of nine dataset. In terms of comparison of classification accuracy using “Wilcoxon signed-rank test,” the finding indicates that “there is no statistically significant differences between the comparative algorithm and the proposed algorithm.” This indicates that KSA could be used as an alternative approach to feature selection for a classification problem.