A Hybrid IPA-ANN Model for Feature Selection and Classification on the Iris Dataset

Abstract

Recently, studies on complex datasets in the field of machine learning (ML) have become widespread. Unnecessary features in these datasets can reduce the accuracy of models and significantly increase computational costs. Therefore, the feature selection (FS) method plays a critical role in determining the most appropriate and effective features according to the model to be designed. Meta-heuristic algorithms are frequently used in this field because they provide effective solutions to FS problems and improve classification performance. Meta-heuristic algorithms provide effective solutions for datasets when FS methods are inadequate. In this study, a new hybrid approach was developed for feature selection and classification on the iris dataset. Feature selection was performed with the maximum separation technique, and then Kernel Principal Component Analysis (KPCA) was used to reduce it to two-dimensional space. The Artificial Neural Network (ANN) model was used for the classification problem. The architecture of the model consists of an input layer with two features, two hidden layers consisting of 64 and 32 units, respectively, and an output layer with three classes. Immune Plasma Algorithm (IPA) was used for hyperparameter optimization. Finally, in order to evaluate the performance of the developed IPA-ANN algorithm in the classification process, accuracy, cross-entropy (CE), F1-Score, and Cohen's Kappa metrics were measured as one-fold, five-fold, and ten-fold. In line with the obtained values, the best values were obtained as 1.0000 for accuracy and 0.0022 for CE in the training part; 1.0000 for accuracy, 0.0003 for CE, 1.0000 for F1-Score, 1.0000 for Cohen's Kappa in the testing part.