4.7 Article

Combinatorial projection pursuit analysis for exploring multivariate chemical data

Journal

ANALYTICA CHIMICA ACTA
Volume 1174, Issue -, Pages -

Publisher

ELSEVIER
DOI: 10.1016/j.aca.2021.338716

Keywords

Exploratory data analysis; Projection pursuit analysis; Cluster analysis; Unsupervised classification; Data visualization; Dimensionality reduction; Grape juice

Funding

  1. Natural Sciences and Engineering Research Council of Canada (NSERC) [46316]
  2. Coordenacao de Aperfeicoamento de Pessoal de Nivel Superior - Brasil (CAPES) [001 (PDSE-CAPES-88881.362191/2019-01)]
  3. Fundacao Araucaria [033/2019]
  4. UTFPR-Campo Mourao (SEI) [23064.010973/2019-26]

Ask authors/readers for more resources

Kurtosis-based projection pursuit analysis (kPPA) is a useful tool for visualizing multivariate data and partitioning binary datasets effectively. The introduction of Combinatorial Projection Pursuit Analysis (CombPPA) allows researchers to explore interesting projections more fully by mapping local minima and visualizing different projections. This approach enhances the orthogonality of solutions obtained and provides insights into various combinations of projections.
Kurtosis-based projection pursuit analysis (kPPA) has demonstrated the ability to visualize multivariate data in a way that complements other exploratory data analysis tools, such as principal components analysis (PCA). It is especially useful for partitioning binary data sets (2k classes) with a balanced design. Since kPPA is not a variance-based method, it can often provide unsupervised class separation where other methods fail. However, when multiple classifications are possible (e.g. by gender, age, disease state, etc.), the projection provided by kPPA (corresponding to the global minimum kurtosis) will not necessarily be the one of greatest interest to the researcher. Fortunately, the optimization algorithm for kPPA allows for interrogation of projections obtained from numerous local minima. This strategy provides the basis of a new method described here, referred to as combinatorial projection pursuit analysis (CombPPA) because it presents alternative combinations of class separation. The method is truly exploratory in that it allows the landscape of interesting projections to be more fully probed. The approach uses Procrustes rotation to map local minima among the kPPA solutions, whereupon the researcher can visualize different projections. To demonstrate the new method, the clustering of grape juice samples using visible spectroscopy is presented as a model problem. This problem is well-suited to this type of study because there are eight classes of samples symmetrically partitioned into two classes by type (organic/non-organic) or four classes by brand. Results presented show the different combinations of projections that can be obtained, including the desired partitions. In addition, this work describes new enhancements to the kPPA algorithm that improve the orthogonality of solutions obtained. (C) 2021 Elsevier B.V. All rights reserved.

Authors

I am an author on this paper
Click your name to claim this paper and add it to your profile.

Reviews

Primary Rating

4.7
Not enough ratings

Secondary Ratings

Novelty
-
Significance
-
Scientific rigor
-
Rate this paper

Recommended

No Data Available
No Data Available