Component Correction and Category Identification for Weathered Silicate Glass With Small Datasets

Machine learning-based models for data classification have found widespread application in scenarios such as industrial fault detection and image recognition. However, identifying the category of weathered silicate glass poses enormous challenges due to the rare quantity, difficulty in sampling, and multiple categories involved. The cost of appraisal instruments for ancient cultural relics is high, and their usage is limited by the environment. For small samples of ancient glass relics, direct application of machine learning algorithms can easily lead to low accuracy and overfitting problems. To address this issue, this article proposes a stacking integration classification and Gaussian mixture clustering (SIC-GMC) model for component correction and category identification of weathered silicate glass with small datasets. The method combines a statistical machine learning model to perform component correction on weathered sampling data to augment the dataset. Next, the stacking integrated learning method based on extreme gradient boosting (XGBoost), random forest, and extra trees is used to identify glass categories. The effectiveness of this method is verified on a practical dataset of ancient glass relics, and in-depth feature mining is performed on the glass category of this instance through a GMC-based subclass division.