A weighted fuzzy C-means clustering method for hardness prediction

The hardness prediction model was established by support vector regression (SVR). In order to avoid exaggerating the contribution of very tiny alloying elements, a weighted fuzzy C-means (WFCM) algorithm was proposed for data clustering using improved Mahalanobis distance based on random forest importance values, which could play a full role of important features and avoid clustering center overlap. The samples were divided into two classes. The top 10 features of each class were selected to form two feature subsets for better performance of the model. The dimension and dispersion of features decreased in such feature subsets. Comparing four machine learning algorithms, SVR had the best performance and was chosen to modeling. The hyper-parameters of the SVR model were optimized by particle swarm optimization. The samples in validation set were classified according to minimum distance of sample to clustering centers, and then the SVR model trained by feature subset of corresponding class was used for prediction. Compared with the feature subset of original data set, the predicted values of model trained by feature subsets of classified samples by WFCM had higher correlation coefficient and lower root mean square error. It indicated that WFCM was an effective method to reduce the dispersion of features and improve the accuracy of model.