Optimization of Hyper Parameter Bandwidth on Naïve Bayes Kernel Density Estimation for the Breast Cancer Classification
Theopilus Bayu Sasongko, Oki Arifin, Hanif Al Fatta · 2019 International Conference on Information and Communications Technology (ICOIACT) · 2019
Cancer is viewed as the most harmful disease in the world. One of the most deadly cancer diseases attacking the women is breast cancer. In this paper, used Naïve Bayes classifier Kernel Density Estimation in breast cancer Wisconsin dataset take from UCI data repository. Naïve Bayes with Kernel Density Estimation was used as it could enable the Naïve Bayes to process the quantitative data. However, the performance of KDE (kernel density estimation) highly depends upon the size of the bandwidth parameter used to control the curve. The problem an accurate choice of parameter bandwidth (h) using trial and error technique requires a long process. A large bandwidth (h) will over-smooth the density and mask the structure in the data, a small bandwidth (h) will yield a density estimate that is spiky and very hard to interpret. The aim of this research is to compare the optimization of grid search parameter with the genetic algorithm to determine the bandwidth parameter (h) on the Naïve Bayes Kernel Density Estimator algorithm. The parameters used to compare the two optimization methods included accuracy, AUC (Area under Curve), and computation time. The experiment results on the test of the significance value on the comparison of grid search and algorithm in determination of the bandwidth parameter (h) naïve Bayes kernel density used Mann-Whitney Test. The result of the significance ranking showed that mean rank of the genetic algorithm accuracy was 489.95 better than the mean value of the rank of grid search accuracy at only 189.05, mean rank of AUC genetic algorithm was 491.82 better than mean rank of AUC grid search only at 187.18. The mean rank of computation time grid search was 243.37 faster than mean rank of computation time genetic algorithm at 435.63.