Enhancing the classification metrics of spectroscopy spectrums using neural network based low dimensional space.

Area of Science:

• Chemometrics
• Machine Learning
• Spectroscopy

Background:

• Spectroscopy quantifies light-matter interactions for analyzing particles, particularly biomolecules.
• Spectral data is high-dimensional, posing challenges for robust classification models.
• Existing dimensionality reduction methods struggle to capture subtle spectral details or handle nonlinearity.

Purpose of the Study:

• To propose a graph-based neural network embedding approach for spectral data dimensionality reduction and classification.
• To overcome the limitations of existing methods in handling spectral data nonlinearity and preserving subtle features.
• To enhance classification performance metrics for spectral datasets.

Main Methods:

• A two-phase dimensionality reduction technique involving nearest neighbor graph construction and fully connected neural network embedding.
• Classification of the low-dimensional embedding using the Random Forest algorithm.
• Comparison with four widely used nonlinear dimensionality reduction techniques on five spectral datasets.

Main Results:

• The proposed approach achieved accuracy scores above 95% and Matthew's correlation coefficient close to 1 across datasets.
• Demonstrated competitive performance across six different low-dimensional spaces for each dataset.
• High trustworthiness scores indicate preservation of the high-dimensional spectral data structure in the latent space.

Conclusions:

• The graph-based neural network embedding approach effectively addresses nonlinearity in spectral data.
• The method offers a robust solution for dimensionality reduction and classification in chemometrics.
• This technique provides a reliable way to extract meaningful features from complex spectral data.