Abstract
The serial genetic algorithms (SGAs) have been widely applied in improving support vector machine (SVM) performance (e.g., classification accuracy), and these hybrid SGA-SVM methods show good capability to detect breast cancer. However, there remain two great challenges: (1) the improvements tend to be at the great cost of time-consuming training; and (2) the SGA-based search may risk the premature convergence to local optima and thereby decrease the quality of the solutions found. The study aimed to investigate the use of parallel genetic algorithms (PGAs) in improving SVM performance, and build an efficient and accurate classifier of detecting breast cancer. A coarse-grained parallel genetic algorithm (CGPGA) was used to select a feature subset and optimize the parameters of SVM simultaneously. This approach (CGPGA-SVM) was then applied to a well characterized breast cancer dataset, consisting of 699 samples (458 benign and 241 malignant samples). In addition, the proposed CGPGA-SVM classier was compared with a range of SVM-based classifiers to understand its performance improvements. Compared with the SGA-SVM classifier, the training time of the CGPGA-SVM classier decreased by 75.77% on a commonly used 4-core CPU; moreover, the classification accuracy and sensitivity of the CGPGA-SVM classifier increased by 0.43% and 1.25%, respectively.