Article Preview
Top1. Introduction
With the development of information in data world, data mining technology is essential to mine the data. The preprocessing task like feature selection is necessary while using data mining tools. FS is one of the best methods extensively used to reduce the dimension of data. The FS scheme protects obtainable feature subsets and removes the rest (supervised or not supervised) during the learning process (Chen et al., 2018). It is defined as the process that enhances the learning algorithm by removing irrelevant and redundant features from the dataset. The FS scheme is categorized into two categories such as strategy and evaluation. After the selection of feature subsets, it can be classified into two techniques as filter and wrapper method. Filter systems select feature subsets based on the data intrinsic properties without concerning any learning process. Hence, the Relief-F and Fisher score supervised methods are included in filter methods (Purushothaman et al., 2020). In wrapper schemes, predictor is considered as a block-box even though the predictor objective functions are evaluating the feature subset. Although these types of methods can achieve good predictive performance but it is time-consuming. Therefore, the problem of dimension reduction is aimed at decreasing the number of features while making the learning process as computationally feasible despite to preserve the power of discrimination between different classes (O’Neill et al., 2018).
A feature selection technique is incorporated into the training process of classifier to minimize the efforts of computation. Not only feature selection, nevertheless selected features weighting or proper scaling will increase the efficiency of the classifier. In high dimensional data, the natural structure may get distorted due to the variations in the scales of dissimilar features. Hence, it is necessary to re-scale the original structure of data (Pal et al., 2021). For determining the optimal weights of the features, the features are selected depending upon the threshold enforced on the weights. On the other hand, the optimal feature selection task is challenging and computationally expensive. In recent times, metaheuristic is effective and reliable for numerous optimization problems in the field of engineering, data mining, machine learning and feature selection. In feature weighting, the weights are assigned according to the discrimination ability of features, which can be used as polynomial, linear or arbitrary functions. In some applications, one must not only determine the relevant features but also provide a priority order in the set of extracted features. Thus, feature weighting can be implemented over the subset of selected features in certain applications (Das & Das, 2017). Moreover, metaheuristic optimization algorithms are used in the data processing applications for better evaluation. In case of FS, different optimization algorithms like Genetic algorithm (GA), Particle Swarm Optimization, Grey Wolf Optimization (GWO), etc., are utilized. The balanced optimal outputs are obtained from these types of metaheuristic optimization algorithms (Dalwinder et al., 2020).
The contribution of this paper is mentioned as follows. A metaheuristic based optimization algorithm is used to perform both feature selection and feature weighting processes. Further, a deep learning network has been employed to perform classification task. At first, suitable features are selected and subsequently allocated feature weights through Tunicate Swarm Genetic Algorithm (TSGA). These features are then used to perform classification by means of DRN classifier. Finally, the performance is evaluated and the superiority is measured in comparison with similar existing approaches.
The rest of the paper is organized as follows. Related techniques on feature selection and feature weighting are reviewed on Section 2. Section 3 and Section 4 explains the proposed methodology and proposed optimization technique respectively. Section 5 discusses the experimental results and finally concluded in Section 6.