Three-stage multi-objective feature selection with distributed ensemble machine and deep learning for processing of complex and large datasets

Abstract

The need for analytical solutions using machine learning (ML) and deep learning (DL) has risen in this era of industry-wide informed decision-making. Data is now readily available and moving at fast speeds due to the maturation of data processing technologies and the development of new technologies in the Big Data, streaming, and IoT sectors. As a result, a transformation is required in how analytical systems handle data. A novel approach to learning, distributed learning, maybe a wise choice for managing huge and high-speed data while overcoming the difficulties presented by noisy and dynamic data sources. Data scientists can considerably increase their productivity using distributed learning by 1) distributing concurrent experiments across several devices and 2) significantly reducing training time by distributing a network's training across several devices. A learning strategy takes a long time, and the model's performance suffers in large datasets due to redundant features and the dimensionality curse. But most feature selection methods are unstable, selecting different subsets of features for different training datasets, resulting in varying levels of classification accuracy. A novel ensemble machine and deep learning technologies are also needed to improve accuracy in distributed learning. To analyze complex and large datasets, this paper proposed a three-stage multi-objective feature selection (TMFS) with distributed ensemble machine and deep learning (DEMDL). The proposed TMFS technique selects the best feature subset by combining multiple feature subsets. To achieve these goals, the TMFS technique employs 5 feature selection strategies in 3 stages: correlation coefficient, Fisher score, information gain, mean absolute deviation, and min-max normalization. SVM is used for machine learning, and Dl4jMlpCassifier is used for deep learning. This SVM and Dl4jMlpClassifier are combined in DEMDL using a stacking classifier. SVM and Dl4jMlpClassifier are utilized as base classifiers in this stacking classifier, and the RIPPER classifier is employed as the meta-classifier. The experimental finding demonstrates that compared to existing classification techniques, the TMFS with DEMDL technique offers higher accuracy, Precision, Recall, and F-measure.