Random Forest Permutation Feature Importance for Feature Selection in Ion Mobility Spectrometry
Abstract
Machine learning (ML) can be used to classify the spectra generated in ion mobility spectrometry (IMS)-based chemical detectors when they are used in the detection of explosives, chemical warfare agents, and other volatile hazardous compounds. The spectra of an IMS detector are composed of drift time bins (x axis) and their corresponding amplitudes (y axis). These values represent the peaks that are present in each spectrum and can be used as features to train ML algorithms. When training on the spectrum, the total number of features can exceed 1000 and are of high cardinality. This work demonstrates a method of using random forest (RF) classification and permutation feature importance to downselect the most important spectral bins. These features are representative of common peak locations in the training data. The use of the most important features determined by RF to reduce the dimensions of the training data set greatly enhanced the accuracy of a dense neural network.
Document Details
- Document Type
- Technical Report
- Publication Date
- Sep 14, 2022
- Accession Number
- AD1180093
Entities
People
- Brian C Hauck
- Brian S. Ince
- Kyle P. O'donnell
- Mary M. Wade
- Patrick C. Riley
- Ruth Dereje
- Samir V. Deshpande