Open Collections

Krasnov, Daniel

Raman spectroscopy is an optical interrogation method capable of providing a unique “fingerprint” for both biological and non-biological compounds. However, due to the multi plexed nature of Raman spectra, machine learning techniques are required to identify which chemical signals are present. This thesis combines group- and basis-restricted non-negative matrix factorization (GBR-NMF) with Bayesian positive source separation (BPSS) to cre ate a probabilistic non-negative matrix factorization technique. Our method not only allows for the incorporation of prior knowledge, but also quantifies the uncertainty in the source separation process, an aspect missing from most standard NMF procedures. The contri butions of this thesis are: a sensitivity analysis to facilitate discussion around prior model selection and constraints, an analysis of Raman spectral data to showcase the efficacy of our approach, and the provision of R code to increase the accessibility of BPSS.

Text

University of British Columbia. DATA 449

Vancouver : University of British Columbia Library

10.14288/1.0443983

Science, Irving K. Barber Faculty of (Okanagan); Computer Science, Mathematics, Physics and Statistics, Department of (Okanagan)

Undergraduate

http://creativecommons.org/licenses/by-nc/4.0/