Evaluating the performance of machine learning and variable selection methods to identify document paper using...

Area of Science:

• Forensic Science
• Analytical Chemistry
• Machine Learning

Background:

• Infrared (IR) spectroscopy offers non-destructive and rapid analysis crucial for forensic examinations.
• Machine learning (ML) and chemometrics are increasingly applied to forensic science, including questioned document examination.
• Identifying document paper manufacturers is vital for forensic investigations.

Purpose of the Study:

• To develop and evaluate ML models for identifying document paper manufacturers using IR spectral data.
• To determine the effectiveness of specific IR spectral regions for model training.
• To assess the performance and computational efficiency of different ML algorithms.

Main Methods:

• Infrared (IR) spectral data from document paper samples were collected.
• Support Vector Machine (SVM), Feedforward Neural Network (FNN), and Random Forest (RF) models were constructed.
• Variable importance from RF models guided the selection of IR spectral regions (1500-800 cm⁻¹).
• Models were trained and evaluated using second-derivative IR spectra within the selected range.

Main Results:

• The FNN and RF models achieved high performance, with F1-scores of 0.978 and 1.000, respectively.
• Selecting specific IR spectral regions significantly enhanced model performance.
• The chosen spectral range minimized computational costs while maintaining high accuracy.
• The models demonstrated robustness in identifying document paper manufacturers.

Conclusions:

• Machine learning methods, particularly FNN and RF, are highly effective for forensic document paper analysis.
• Optimizing spectral data selection enhances ML model performance and efficiency in forensic applications.
• This approach provides a robust and computationally inexpensive tool for identifying document paper origins.