br Experimental design br This section presents the
6.3. Experimental design
This section presents the details of our experimental design. In order to verify the effective performances of our proposed model, we design the corresponding comparative experiment from two aspects. In the first aspect, we will compare the performances of our proposed IGSAGAW feature selection approach with GAW method, as can be seen from Table 7, it BAY-598 presents the main parameters of SAGAW algorithm. Additionally, in order to emphasis the superiority performances of feature selection, we ran the experiment on the
Rang of each attributes in WDBC data set.
Attribute Mean Standard error Maximum
Predicted positive (benign) Predicted negative (malignant)
Actual positive (benign) True positive(TP) False negative(FN) Actual negative (malignant) False positive(FP) True negative (TN)
The main parameters of SAGAW algorithm.
Maximum number of generation 200 The size of the population 50 Selection type Tournament selection Cross type Single point cross Mutation Rate 0.1 Mutation Type Uniform mutation Initial temperature value 100 Temperature decay coefficient 0.9
basis of underlying classifiers with all features, then applying different feature selection approaches of GAW and IGSAGAW using three typical classifiers.
The aim of our second aspect is to explore the effect of the different underlying classifiers. To the best of our knowledge, in the field of machine learning, breast cancer diagnosis has been considered as classification problem, and the classification approaches such as BP neural network, K-NN, CSSVM were considered as excellent classifiers, which have been utilized as underlying classifiers in our experimental. And the parameters of different comparative methods are presented in Table 8. In this work, we adopt grid search approach for finding the optimal parameters of SVM and set the search range of parameter C = 2 10 , 2 8 , …, 28 , 210 and
g = 2 10 , 2 8 , …, 28 , 210 , respectively, then the value of C, g searching step is set to 0.5. The main objective of parameter searching is to find the best parameter pair of (C, g), which can achieve the best performances. After obtaining the best parameter pair, we create the classification model performing for training and testing. To ensure the rationality of our experimental results, we utilized the data set which processed by the feature selection method as described above, and adopt 10-fold cross validation for WBC and WDBC data sets. In each comparative approach, we take into account the different feature selection methods as described above and design the source codes, which were implemented in MATLAB platform. In order to eliminate the randomness factor and reflect the results rationality, we employed 10-fold cross verification and the average performance of these results were reported as the final results in Lytic infection study.
6.4. Experimental results and analysis
In our experimental, we design the source code of our proposed algorithm, which implemented on MATLAB platform. The results of 10-fold cross validation on different performances as can be observed in Figs. 2–9. As previous mentioned, in order to straight-forward assess the effectiveness of our proposed model, we carried out a series of comparative experiments, and the experimental results analysis as described below.
6.4.1. Classification accuracy for the best solution
First of all, in order to verify the effect of our proposed model, we first ran the experiments on the baseline classifiers with all the features before applying our IGSAGAW and GAW feature selection approaches. As can be seen from Figs. 2 and 3, the average classification accuracy of 10-fold cross verification on IGSAGAW is higher than GAW, and the accuracy of GAW is higher than baseline classifier. The main reason is that in GAW feature selection approach, we select the top n features which obtained the best value of fitness. And during this feature selection process, we applied BP, 3-NN and CS-SVM three underlying classifiers perform for classification. In IGSAGAW approach, we utilized IG ranking the importance of features firstly, then we applied SAGAW algorithm to