A Robust TabNet-Based Multi-Classification Algorithm for Infrared Spectral Data of Chinese Herbal Medicine with...

Area of Science:

• Chemometrics
• Machine Learning
• Spectroscopy

Background:

• High-dimensional, small-sample datasets pose challenges for classification algorithms.
• Accurate identification of Chinese medicinal materials is crucial for quality control and efficacy.

Purpose of the Study:

• To develop a novel embedded multiclassification algorithm, ITabNet, for analyzing infrared spectroscopic data of Chinese medicinal materials.
• To address challenges of high dimensionality and small sample size in classification tasks.

Main Methods:

• ITabNet, derived from TabNet, incorporates a refined data pre-processing (DP) mechanism optimized using Support Vector Machine (SVM).
• An innovative focal loss function and cross-validation strategy were employed to handle sample imbalance.
• ITabNet performance was compared against SVM and Extreme Gradient Boosting (XGBT) under various conditions (DP/Non-DP, GPU/CPU).

Main Results:

• ITabNet significantly improved prediction effectiveness, achieving perfect accuracy and Area Under the Curve (AUC) scores of 1.0000.
• The algorithm demonstrated superior performance compared to SVM and XGBT.
• The refined DP mechanism efficiently identified optimal pre-processing methods.

Conclusions:

• ITabNet is a robust and effective algorithm for classifying high-dimensional, small-sample datasets, particularly in the context of infrared spectroscopy for Chinese medicinal materials.
• The study provides valuable insights for multi-classification modeling with limited sample sizes and numerous features.
• ITabNet shows potential for analyzing medicinal efficacy and chemical composition.