Abstract:

Near-infrared spectroscopy technology is highly effective in pharmaceutical analysis. For high-dimensional and non-linear small-scale near-infrared data, traditional drug identification algorithms lack enough feature learning ability, neural network-based methods have problems of local optima and over-fitting, and they tend to ignore the sample imbalance. Aiming at the above disadvantages, a pharmaceutical classification approach with multi-layered gradient Boosting decision trees based on feature selection and cost-sensitive learning (CS_FGBDT) is proposed. Firstly, the raw data are preprocessed by Savitsky-Golay smoothing and first derivative. Secondly, the random forest is used to adaptively extract features from the preprocessed spectra, and the feature map is constructed by multi-layered gradient Boosting trees. Then the negative effect of sample imbalance is minimized by combining cost-sensitive learning. The experimental results show that the model comparatively evaluated on two imbalanced data-sets of capsule and tablet has higher prediction accuracy and stability and is an effective method for drug identification.

Key words: near-infrared spectroscopy, adaptive feature selection, multi-layered gradient Boosting decision trees, cost-sensitive learning

摘要：

近红外光谱分析技术高效应用于药品分析领域。针对高维非线性的小规模近红外数据，传统的药品鉴别算法存在特征学习能力不足的缺陷，基于神经网络的方法有局部最优及过拟合等问题，且两者易忽略样本的不均衡性。针对以上劣势，提出一种基于特征选择与代价敏感学习的多层梯度提升树（CS_FGBDT）药品分类方法。首先采用Savitsky-Golay平滑和一阶导数对原始数据进行预处理；其次利用随机森林对预处理光谱自适应提取特征，并由多层梯度提升树进行特征映射；然后结合代价敏感学习机制将样本不均衡性的负效应降到最小。实验结果表明，在胶囊和药片两种不平衡数据集上对算法进行对比评估，该模型具有更高的预测精度和稳定性，是一种有效的药品鉴别方法。

关键词: 近红外光谱分析, 自适应特征选择, 多层梯度提升决策树, 代价敏感学习