Ardışıl ileri yönlü öznitelik seçim algoritmasında etkin özniteliklerin belirlenmesi

In recent years, pattern recognition and machine learning has become a very active research area due to the applicability in various kind of subjects, including brain computer interface, commercial and financial approaches. Furthermore, it is worthwhile mentioning that there is no definite technique to solve any classification problem. Thus, it can be helpful to get the most discriminative features using feature selection algorithm. Such approaches are generally performed in four steps: 1- preprocessing, 2- feature extraction, 3- feature selection and 4- classification/regression. Among these steps, feature extraction and feature selection are very vital for representing input signals in a reduced feature space and for identifying discriminative information in order to propose fast and high performance application. Because of there are many kind of feature extraction method, in some cases hundreds of features might be calculated for identifying input signals. However, enormous number of features might reduce the decision performance and speed which are very crucial two parameters in pattern recognition and machine learning approaches. In order to eliminate those disadvantages and reduce the number of features some techniques have been proposed. Those techniques are used by machine learning community for selecting the most suitable feature subset among all the extracted feature sets in order to increase the performance of the proposed model. The sequential forward feature selection (SFFS) and the sequential backward feature selection (SBFS) algorithms are very widely used feature selection techniques in literature. In this study, we extracted Continuous Wavelet Transform based features from BCI Competition 2005 Data Set I. Afterwards, among the extracted features, the most stable and effective features were selected by SFFS and SBFS techniques. BCI Competition 2005 Data Set I includes electrocorticogram (ECoG) based brain computer interface signals which was taken from an epilepsy subject on two different days with about one week of delay. In the both sessions the ECoG signals were recorded while subject was asked to imagine of either the left small finger or the tongue movement. The signals were acquired with an 8x8 ECoG platinum electrode grid (totally from 64 points) which was placed on the contralateral (right) motor cortex. All recordings were performed with a sampling rate of 1 kHz (acquired 3000 samples per channel for every trial). Additionally, BCI Competition 2005 Data Set I consist of 278 training trials (139 trials for finger movements, 139 trials for tongue movements) and 100 test trials which were recorded in the first session and the second session, respectively. Each trial’s duration was 3 seconds. In this paper, SFFS and SBFS algorithms tested for determining of effective features after feature extraction procedure. Afterwards, they compared in terms of classification accuracy and speed. While those methods determine the effective features from training data set using cross validation method, the sub-training data sets are selected randomly. So that, different features might be selected for every running of those methods. Thus, selecting different features are influenced the test performance of the model positively/negatively, as well. Moreover, in this paper a method is proposed to overcome this randomly selection disadvantage. In order to show the robustness of the proposed method the SFFS and SBFS algorithms were run 1000 times in the training stage. Afterwards the features, which were selected more than the determining threshold level, were selected as effective features. Moreover, SFFS and SBFS algorithms were compared in terms of the speed and classification accuracy. The obtained results showed that, SFFS is approximately 40 times faster than SBFS and SFFS provides more than 22% classification accuracy.