4:30 PM - 4:45 PM
[MGI28-05] Utility of geochemical data analysis based on non-Gaussianity
★Invited Papers
Keywords:non-Gaussianity, independent component analysis, unsupervised clustering, projection pursuit, multivariate analysis, geochemical data analysis
The advancement of analytical chemistry since the 20th century enabled us to obtain multi-elemental information from various geological samples very efficiently. As a consequence, the accumulation of geochemical data has continued to increase. In recent years, along with the dramatic growth in the processing capabilities of computers, data-driven approaches to extract useful information from large-scale/high-dimensional geochemical data have proceeded vigorously. The ultimate goal of these efforts is to unravel the information about the origin of various samples obtained from nature and the physicochemical processes (or environment) involved in their formation. This type of problem is known as the "Blind Source Separation" (BSS) in signal processing, which is the problem of recovering multiple source signals mixed in unknown proportions using only the observational data available to us as "results" in the natural world. In other words, the essence of geochemical data analysis can be described as solving the BSS problem.
In this presentation, as one of the solutions to the BSS problem, we will explain the concept, principles, and algorithms of Independent Component Analysis (ICA). ICA can reveal the hidden structures in high-dimensional data, focusing on the non-Gaussianity inherent in the data [1]. We will introduce the results obtained by the application of ICA to various geochemical data [2-4]. Furthermore, the ICA algorithm is closely related to a method called "projection pursuit," which is used in unsupervised clustering to produce visualizations that best capture the characteristics of data structure. We will also touch upon our recent work obtained by applying this method (construction of high-dimensional chemostratigraphy inherent in pelagic clays [5]). Finally, we will discuss the relationship between ICA and the constant-sum constraint problem that is an unresolved issue in compositional data analysis, as well as the limitations of this method.
[1] Hyvärinen et al. (2001) Independent Component Analysis. [2] Yasukawa et al. (2016) Scientific Reports 6, 29603. [3] Yasukawa et al. (2017) Scientific Reports 7, 11304. [4] Yasukawa et al. (2022) Chemical Geology 614, 121184. [5] Yasukawa et al. (2023) Paleoceanography and Paleoclimatology 38, e2023PA004644.
In this presentation, as one of the solutions to the BSS problem, we will explain the concept, principles, and algorithms of Independent Component Analysis (ICA). ICA can reveal the hidden structures in high-dimensional data, focusing on the non-Gaussianity inherent in the data [1]. We will introduce the results obtained by the application of ICA to various geochemical data [2-4]. Furthermore, the ICA algorithm is closely related to a method called "projection pursuit," which is used in unsupervised clustering to produce visualizations that best capture the characteristics of data structure. We will also touch upon our recent work obtained by applying this method (construction of high-dimensional chemostratigraphy inherent in pelagic clays [5]). Finally, we will discuss the relationship between ICA and the constant-sum constraint problem that is an unresolved issue in compositional data analysis, as well as the limitations of this method.
[1] Hyvärinen et al. (2001) Independent Component Analysis. [2] Yasukawa et al. (2016) Scientific Reports 6, 29603. [3] Yasukawa et al. (2017) Scientific Reports 7, 11304. [4] Yasukawa et al. (2022) Chemical Geology 614, 121184. [5] Yasukawa et al. (2023) Paleoceanography and Paleoclimatology 38, e2023PA004644.