Improved CNN Based Feature Extraction for Data Mining

Abstract

Data mining extricates the required data from the cluster of information's. Based on the given concept the irrelevant data are mined from the featured information. The proposed method includes four phases such as preprocessing, oversampling, feature extraction and classification. Initially, the dataset is preprocessed in three steps namely cleaning, encoding and normalization. In this paper, Spam base, WaveformEW, Wisconsin Diagnostic Breast Cancer (WDBC), SonarEW and SPECTEW datasets are utilized but which are not properly balanced and creates class imbalance problem. Therefore, Adaptive Synthetic(ADASYN) technique is utilized to oversampling the minority class. Feature extraction is done by animproved Convolution Neural Network (CNN). CNN is improved by clipped leaky rectified linear unit, rectified linear unit and Rectified linear unit (ReLU)activation functions for feature extraction. After feature extraction, SVM classifier is applied to classify the features and then it rectifies the class imbalance issue. The proposed method is evaluated with five datasets fromUCI in terms of precision, accuracy, specificity, sensitivity, etc. Based on the analysis the WDBC dataset provide better results in CNN with Clipped ReLU method. To compare with ReLU and Leaky ReLU, the Clipped ReLU method achieves higher accuracy of 99.30%.