Chemical Composition Identification Model Based on a Variety of Unsupervised Machine Learning Classification Methods

This paper begins by visually representing the data and examining the classification patterns between highpotassium glass and lead-barium glass. Subsequently, we employ SPSSPRO for K-means clustering, utilizing the elbow method to ascertain the optimal value of $\mathbf{K}$. Post-classification, we further subclassify the characteristics of the cluster center coordinates for each index within the sample dataset. By establishing R-type clustering to subclassify the data again, comparing the results of K -means clustering and R -type clustering to judge the sensitivity and stability of the classification model, it is concluded that K-means clustering for subclassification is reasonable and clear, and has good stability and consistency. Then, the data were preprocessed using the isometric logarithmic ratio transformation, considering that the “closure effect” caused by the numerical values would lead to pseudo-correlation of the chemical data. Next, PCA downscaling was performed on the processed data to reduce the data to 2 dimensions. Finally, Form 3 glass categories were classified according to this classification threshold. Use decision tree classification to reclassify the attached Form 3, and then compare the results of the decision tree classification model with the results of the principal component threshold classification model, and conclude that the classification model is less sensitive to parameters and has better stability.