In recent years, a lot of data in different fields has been recorded by increasing the development in databases and information systems, as well as using them in various scientific areas for modeling and decision-making purposes. Computational cost and time are increasingly increased if all of the provided data from the underlying systems are simultaneously used. One of the most reasonable and efficient techniques in such situations is to only use the most important and effective data and variables. In the literature, several different approaches have been proposed that can be generally categorized into two main categories of feature extraction and feature selection. In feature extraction approaches, all variables are preserved; nevertheless, they are converted to a smaller space by a linear or nonlinear conversion function. While in the feature selection class, the space of variables is preserved; however, the number of variables is reduced. Regression analysis is one of the most widely-used methods in the statistical and linear data analysis domain in which dimensional reduction is frequently used. Forward and backward dimensional reduction methods are among the most popular approaches in regression analysis. However, these methods, despite all their unique advantages, have some critical disadvantages that may decrease their performance in some specific situations. First, in these methods, a criterion is only used to prioritize variables for adding and/or removing, in which the relationship between independent variables and the dependent variable is only considered. Second, the ordinary correlation coefficient, in which the risk of losing appropriate variables is high, is applied. In this thesis, a new feature selection method is proposed in order to overcome the aforementioned limitations of the forward as well as backward methods. In the proposed model, first, two criteria are simultaneously considered for prioritization and eva‎luation of entering and/or exiting variables in the forward and backward processes. In this way, the correlation matrix is first used to hierarchically cluster variables, and then the partial correlation coefficient is used for determining the priority of variables and clusters. Three feature selection benchmarks are considered in order to eva‎luate the performance of the proposed model in comparison with the conventional forward and backward methods. Empirical results indicate that the proposed method can yield more accurate results than conventional forward and backward methods in all case studies.

مهدي خاشعي

علي زينل همداني، فرشته پرورش