Machine Learning Driven Feature Sensitive Progressive Sampling Model for BigData Analytics

Abstract

BigData requires processing a huge data volume, which is an undeniable challenge for academia-industries. The classical sampling techniques are limited when addressing data-imbalance, large data-heterogeneity, multi-dimensionality etc. To alleviate it, in this paper a novel machine learning driven feature sensitive progressive sampling (ML-FSPS) that in conjunction with an improved feature selection and classification environment achieves more than 95.7% of accuracy, even with 10-14% of the original data size. The proposed ML-FSPS model was applied for IoT-device classification problem that possesses exceedingly high data-imbalance, multi-dimensionality and heterogeneity issues. Functionally, the FSPS-driven analytics model at first performed active period segmentation followed by multi-dimensional (descriptive) statistical feature extraction and Wilcoxon Rank Sum Test based feature selection. Subsequently, it executed K-Means clustering over a gigantically huge feature instances (network traces) Here, K-means algorithm clustered each feature samples into five distinct clusters. With initial sample size of 10%, FSPS model selected same amount of data elements (0.5-5% iteratively) from each cluster for each feature to perform multi-class classification using homogenous ensemble learning (HEL) model. Here HEL encompassed AdaBoost, Random Forest and Extended Tree ensemble algorithms as base classifiers. The simulation results affirmed that the proposed model achieves accuracy of almost 99% even with 10-16% of sample size.