An Unsupervised Based Stochastic Parallel Gradient Descent For Fcm Learning Algorithm With Feature Selection For Big Data

Abstract

Huge amount of the dataset consists millions of explanation and thousands, hundreds of features, which straightforwardly carry their amount of terabytes level. Selection of these hundreds of features for computer visualization and medical imaging applications problems is solved by using learning algorithm in data mining methods such as clustering, classification and feature selection methods .Among them all of data mining algorithm clustering methods which efficiently group similar features and unsimilar features are grouped as one cluster ,in this paper present a novel unsupervised cluster learning methods for feature selection of big dataset samples. The proposed unsupervised cluster learning methods removing irrelevant and unimportant features through the FCM objective function. The performance of proposed unsupervised FCM learning algorithm is robustly precious via the initial centroid values and fuzzification parameter (m). Therefore, the selection of initial centroid for cluster is very important to improve feature selection results for big dataset samples. To carry out this process, propose a novel Stochastic Parallel Gradient Descent (SPGD) method to select initial centroid of clusters for FCM is automatically to speed up process to group similar features and improve the quality of the cluster. So the proposed clustering method is named as SPFCM clustering, where the fuzzification parameter (m) for cluster is optimized using Hybrid Particle Swarm with Genetic (HPSG) algorithm. The algorithm selects features by calculation of distance value between two feature samples via kernel learning for big dataset samples via unsupervised learning and is especially easy to apply. Experimentation work of the proposed SPFCM and existing clustering methods is experimented in UCI machine learning larger dataset samples, it shows that the proposed SPFCM clustering methods produces higher feature selection results when compare to existing feature selection clustering algorithms , and being computationally extremely well-organized.
DOI: 10.17762/ijritcc2321-8169.150727