Polynomials feature-transformed heap dimensionality reduction and stacking ensemble for spectrometry data classification
Pattern recognition has emerged as a burgeoning field of study with increasing prominence in light of technological advancements, finding applications across various multidisciplinary domains. An essential part of pattern recognition is classification where it involves the categorization of labelled...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English English |
| Published: |
2023
|
| Subjects: | |
| Online Access: | https://eprints.ums.edu.my/id/eprint/40556/1/24%20PAGES.pdf https://eprints.ums.edu.my/id/eprint/40556/2/FULLTEXT.pdf https://eprints.ums.edu.my/id/eprint/40556/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Pattern recognition has emerged as a burgeoning field of study with increasing prominence in light of technological advancements, finding applications across various multidisciplinary domains. An essential part of pattern recognition is classification where it involves the categorization of labelled samples based on their data features. Fourier Transform Infrared (FTIR) spectroscopy, a well-established spectroscopic technique, have long been used to detect organic, polymeric, and even inorganic materials. This research endeavours to develop an accurate and optimal classification framework on FTIR spectra data using a combination of heap dimensionality reduction (DR) technique, polynomial features transformation and a heuristic stacking ensemble technique. The high-dimensionality nature of FTIR data poses a significant challenge for classification. To address this issue, DR techniques are used. However, no DR technique is superior to all others. Depending on the dataset used, one method may produce a better approximation of a dataset than the other techniques. In this study, the high-dimensional data undergo multiple existing DR techniques. The resulting transformed features are consolidated into a heap and subsequently undergo polynomial feature transformation. Then Partial Least Square (PLS-DA) method is applied to obtain the final transformed features. The transformed features are then utilized as input for the stacking ensemble (SE) model, selected through a heuristic SE procedure. Artificial data was employed for the initial two experiments, while the complete framework was tested on the six FTIR datasets for the third experiment to assess its applicability to real-world datasets. The experimental results on these six datasets revealed that the proposed framework was outperformed the other examined models. Notably, an average accuracy, sensitivity, and specificity of up to 99% was achieved for the D06 dataset. As a result, this framework holds potential not only for the classification of FTIR data but also for other high-dimensional data in general. |
|---|