UTKARSH KHAIRE
@iiitdwd.ac.in
Assistant Professor- Dept. of Data Science and Intelligent Systems
INDIAN INSTITUTE OF INFPRMATION TECHNOLOGY DHARWAD
RESEARCH INTERESTS
Machine learning, Deep learning, Feature selection, and Optimization Techniques
Scopus Publications
Scholar Citations
Scholar h-index
Scholar i10-index
Scopus Publications
Utkarsh Mahadeo Khaire and R. Dhanalakshmi
Elsevier BV
Abstract Feature selection technique is a knowledge discovery tool which provides an understanding of the problem through the analysis of the most relevant features. Feature selection aims at building better classifier by listing significant features which also helps in reducing computational overload. Due to existing high throughput technologies and their recent advancements are resulting in high dimensional data due to which feature selection is being treated as handy and mandatory in such datasets. This actually questions the interpretability and stability of traditional feature selection algorithms. The high correlation in features frequently produces multiple equally optimal signatures, which makes traditional feature selection method unstable and thus leading to instability which reduces the confidence of selected features. Stability is the robustness of the feature preferences it produces to perturbation of training samples. Stability indicates the reproducibility power of the feature selection method. High stability of the feature selection algorithm is equally important as the high classification accuracy when evaluating feature selection performance. In this paper, we provide an overview of feature selection techniques and instability of the feature selection algorithm. We also present some of the solutions which can handle the different source of instability.
Kulanthaivel Balakrishnan, Ramasamy Dhanalakshmi, and Utkarsh Khaire
Wiley
Utkarsh Mahadeo Khaire, R. Dhanalakshmi, K. Balakrishnan, and M. Akila
World Scientific Pub Co Pte Ltd
The aim of this research critique is to propose a hybrid combination of Opposition-Based Learning and Sailfish Optimization strategy to recognize the salient features from a high-dimensional dataset. The Sailfish Optimization is a swarm-based metaheuristics optimization algorithm inspired by the foraging strategy of a group of Sailfish. Sailfish Optimization explores the search space in only one direction, limiting its converging capacity and causing local minima stagnation. Convergence will be optimal if the search space is reconnoitred in both directions, improving classification accuracy. As a result, combining the Opposition-Based Learning and Sailfish Optimization strategies improves SFO’s exploration capability by patrolling the search space in all directions. Sailfish Optimization Algorithm based on Opposition-Based Learning successfully amalgamates the model to global optima at a faster convergence rate and better classification accuracy. The recommended method is tested with six different cancer microarray datasets for two different classifiers: the Support Vector Machine classifier and the K-Nearest Neighbor classifier. From the results obtained, the proposed model aided with Support Vector Machine outperforms the existing Sailfish Optimization with or without K-Nearest Neighbor in terms of convergence capability, classification accuracy, and selection of the most delicate salient features from the dataset.
Utkarsh Mahadeo Khaire, R. Dhanalakshmi, and K. Balakrishnan
CRC Press
K. Balakrishnan, R. Dhanalakshmi, and Utkarsh Mahadeo Khaire
Springer Science and Business Media LLC
Kulanthaivel Balakrishnan, Ramasamy Dhanalakshmi, and Utkarsh Khaire
Wiley
Utkarsh Mahadeo Khaire and R. Dhanalakshmi
Informa UK Limited
Utkarsh Mahadeo Khaire and R Dhanalakshmi
Oxford University Press (OUP)
Abstract A microarray dataset contains thousands of DNA spots covering almost every gene in the genome. Microarray-based gene expression helps with the diagnosis, prognosis and treatment of cancer. The nature of diseases frequently changes, which in turn generates a considerable volume of data. The main drawback of microarray data is the curse of dimensionality. It hinders useful information and leads to computational instability. The main objective of feature selection is to extract and remove insignificant and irrelevant features to determine the informative genes that cause cancer. Random forest is a well-suited classification algorithm for microarray data. To enhance the importance of the variables, we proposed out-of-bag (OOB) cases in every tree of the forest to count the number of votes for the exact class. The incorporation of random permutation in the variables of these OOB cases enables us to select the crucial features from high-dimensional microarray data. In this study, we analyze the effects of various random forest parameters on the selection procedure. ‘Variable drop fraction’ regulates the forest construction. The higher variable drop fraction value efficiently decreases the dimensionality of the microarray data. Forest built with 800 trees chooses fewer important features under any variable drop fraction value that reduces microarray data dimensionality.
K. Balakrishnan, R. Dhanalakshmi, and Utkarsh Mahadeo Khaire
Springer Science and Business Media LLC
Utkarsh Mahadeo Khaire and R. Dhanalakshmi
Springer Science and Business Media LLC
Classifying data samples into their respective categories is a challenging task, especially when the dataset has more features and only a few samples. A robust model is essential for the accurate classification of data samples. The logistic sigmoid model is one of the simplest model for binary classification. Among the various optimization techniques of the sigmoid function, Adam optimization technique iteratively updates network weights based on training data. Traditional Adam optimizer fails to converge model within certain epochs when the initial values for parameters are situated at the gentle region of the error surface. The continuous movement of the convergence curve in the direction of history can overshoot the goal and oscillate back and forth incessantly before converging to the global minima. The traditional Adam optimizer with a higher learning rate collapses after several epochs for the high-dimensional dataset. The proposed Improved Adam (iAdam) technique is a combination of the look-ahead mechanism and adaptive learning rate for each parameter. It improves the momentum of traditional Adam by evaluating the gradient after applying the current velocity. iAdam also acts as the correction factor to the momentum of Adam. Further, it works efficiently for the high-dimensional dataset and converges considerably to the smallest error within the specified epochs even at higher learning rates. The proposed technique is compared with several traditional methods which demonstrates that iAdam is suitable for the classification of high-dimensional data and it also prevents the model from overfitting by effectively handling bias-variance trade-offs.
Utkarsh Mahadeo Khaire and R. Dhanalakshmi
Inderscience Publishers
Utkarsh Mahadeo Khaire and R. Dhanalakshmi
IEEE
Selection of important feature from the high dimensional dataset is a very important task. Irrelevant and insignificant features can hinder the important information of the dataset. For accurate classification of the dataset, selection of most predictive variable is a much-needed task. Feature selection is important for the diagnosis, prognosis and treatment of any disease in case of healthcare dataset. Traditional feature selection algorithm gives the output as a single feature subset of the predictive variable. But the selection of a single feature subset cannot give the perfect idea about the nature of the dataset. The study shows that variable which is not selected by some of the feature selection algorithms will also play a major role in defining the target value of the data sample. Statistically Equivalent Signature (SES) algorithm even consider the set of such important features by considering the Markov blanket of the target variable. The root of SES is in the causal theory and Bayesian network. SES select subset of important features based on the concept of conditional independence. Finally a subset of equally predictive features/variables selected based on the Markov Blanket of target variable T. This leads to improved accuracy when considered with other feature selection algorithms. This paper presents a detailed study for selecting multiple predictive feature subsets that are statistically equivalent and their comparison with the existing feature selection algorithms. It gives an overall idea to process high dimensional datasets especially microarray (gene expression) which play a vital role in the estimation of deadly diseases such as cancer.