成分数据
遥感
环境科学
组分(热力学)
期限(时间)
计算机科学
地质学
数据挖掘
植被(病理学)
领域(数学)
作者
Khaled Masoumifard,Stephan van der Westhuizen,Sugnet Gardner‐Lubbe
出处
期刊:Chapman & Hall/CRC eBooks
[Informa]
日期:2026-04-30
卷期号:: 66-81
被引量:1
标识
DOI:10.1201/9781003564423-4
摘要
Compositions are observations that are unit-sum constrained and frequently encountered in environmental studies, such as species abundance in ecological research, the relative abundance of greenhouse gases in climate studies, and soil texture in soil science. The primary goal of compositional data analysis is to account for the composite nature of observations and to ensure that model outputs adhere to the same constraints as the original data. However, conventional machine learning models may not optimally handle compositional data and may therefore require additional transformations, such as the isometric log-ratio transformation. Another modelling approach involves utilising Dirichlet regression to model compositional data directly in the simplex space. Nevertheless, this approach, like many other parametric models, may not typically achieve the predictive accuracy associated with machine learning models. Consequently, there is a need to explore machine learning approaches specifically designed for compositional data. In this research, we investigate the use of a Dirichlet-adapted random forest capable of training directly in the simplex space. In a classical random forest model, trees are split based on the mean-squared-error (in regression). We propose splitting nodes by maximising the log-likelihood of the Dirichlet distribution. This approach extends the beta-random forest, which is used for modelling bounded outcomes. The Dirichlet-random forest ensures that the resulting predictions maintain the unit-sum constraint while achieving competitive predictive accuracy.
科研通智能强力驱动
Strongly Powered by AbleSci AI