Collinear datasets augmentation using Procrustes validation sets.

Area of Science:

• Machine Learning
• Chemometrics
• Data Science

Background:

• High-complexity models like artificial neural networks (ANNs) require large datasets to prevent overfitting and ensure reproducibility.
• Experimental datasets, especially spectroscopic data, are often limited in size and exhibit high collinearity.
• Existing data augmentation methods struggle with collinearity or are computationally expensive.

Purpose of the Study:

• To introduce an efficient and scalable data augmentation method for collinear datasets.
• To enhance the performance of regression and classification models using augmented data.
• To address the limitations of current augmentation techniques for spectroscopic and similar data.

Main Methods:

• A novel data augmentation approach combining latent variable modeling and cross-validation resampling.
• Application to datasets with moderate to high collinearity, focusing on spectroscopic data.
• Validation using artificial neural networks for prediction and classification tasks.

Main Results:

• Significant improvements in artificial neural network model performance for both prediction and classification tasks.
• Demonstrated effectiveness in case studies involving fat content prediction in minced meat and olive discrimination using near-infrared spectra.
• Achieved up to a 3-fold reduction in root mean squared error for fat content prediction on an independent test set.

Conclusions:

• The proposed method offers a fast, simple, and versatile solution for augmenting collinear datasets.
• It significantly enhances model performance without complex parameter tuning.
• Provides a practical alternative to existing, more resource-intensive data augmentation techniques.