Spark-based deep classifier framework for imbalanced data classification

Abstract

ABSTRACT Imbalanced data are those datasets that have an uneven distribution of observations across the target class; one class label has a very high number of observations while the other has a very low number. The data imbalance occurs when the data are biased towards a certain class and exclude the minority class, which has major significance in the accomplishment of the classifier. The data imbalance problem provides a consequence of poor predictive performance and to avoid this, multiple researchers are engaged in developing multiclass classification techniques for handling imbalance issues. In this research, the data imbalance problem is conquered using Synthetic Minority Oversampling technique (SMOTE) technique that equally distributes the data. The multiclass classification is boosted by employing the Spark architecture and the classification is executed utilising the Lupusbug-deep Convolutional Neural Network (Lupusbug optimisation-deep CNN) classifier that effectively classifies the data through finding the optimal weights of the classifier. Importantly, incremental learning is initiated in this research based on the error function and the selection criterion that facilitates the classifier remodelling and improves the tendency to deal with the incremental data. The proposed method attains 94.348% accuracy, 93.510% precision and 96.121% recall, which was quite efficient and outperforms the existing methods.