基于太赫兹时域光谱数据的柴胡鉴别多分类器比较
Comparison of Multiple Classifiers for Bupleurum Identification Based on Terahertz Time-Domain Spectroscopic

随着机器学习领域的发展，研究人员不断探索新的分类算法模型，使得可供选择的机器学习算法种类更加丰富。然而，许多研究仅使用有限的分类算法，这导致综合比较分类器性能变得困难。为此，本实验利用柴胡太赫兹(THz)时域光谱数据，使用多个评价指标，评估了支持向量机(SVM)、KNN、决策树(Decision Tree, DT)、随机森林(Random Forest, RF)、Logistic回归(LR)、多层感知(MLP)、伯努利朴素贝叶斯(Bernoulli Naive Bayes, BNB)、AdaBoosting、梯度提升决策树(Gradient Boosting Decision Tree, GBDT)、极端随机树(Extremely Random Forest, ERF)、极致梯度提升(eXtreme Gra-dient Boosting, XGB)和轻量梯度提升机(Light Gradient Boosting Machine, LGBM)等12种分类器的分类性能。结果表明，LR、MLP、SVM和KNN分类效果最好，其中，MLP的批次内投票准确率达100%，且召回率和F2得分都较为优异；此外，GBDT、AdaBoosting和LGBM等算法的柴胡鉴别准确度也普遍超过80%。本文为基于THz的柴胡鉴中的分类器选择提供了重要参考。
With the development of machine learning, researchers are constantly exploring new classification algorithm models, making the variety of machine learning algorithms available more diverse. However, many studies only use limited classification algorithms, which makes it difficult to comprehensively compare the performance of classifiers. For this purpose, this paper used terahertz (THz) time-domain spectral data of Bupleurum to evaluate the performance of 12 classifiers including Support vector machine (SVM), KNN, Decision Tree (DT), Random Forest (RF), Logistic Regression (LR), Multilayer Perceptron (MLP), Bernoulli Naive Bayes (BNB), AdaBoosting, Gradient Boosting Decision Tree (GBDT), Extremely Random Forest (ERF), eXtreme Gradient Boosting (XGB) and Light Gradient Boosting Machine (LGBM), in terms of multiple classification performance indicators. The results showed that LR, MLP, SVM, and KNN are the four classifiers with the best classifi-cation performance. Among them, the MLP classifier reaches 100% accuracy after voting and has superior recall and F2 score; in addition, newer algorithms such as GBDT, AdaBoosting and LGBM have also been generally found to have accuracies of more than 80%. This paper provides an im-portant reference for practical applications in the field of Chai Hu identification based on THz.