An Exploration of Projection Pursuit Analysis and Independent Component Analysis as Alternatives for the Analysis of Multivariate Chemical Data
This work presents a critical evaluation of key algorithms for PPA and ICA in applications related to clustering and signal extraction for chemical data. The relationship between PPA and ICA, which has been alluded to in the literature, is firmly established through theory and application. It is demonstrated through application to selected data sets that, while useful for certain kinds of problems, these methods are likely to have limited utility for signal extraction in chemistry, where the source signals are rarely statistically independent and have unpredictable distributions. In contrast, both methods are shown to be useful for clustering applications, but PPA is generally more powerful because of its ability to explore the full variable space. This is demonstrated through a case study where traditional exploratory data analysis methods fail due to a complex error structure in the data.