Understanding complex analytical data by a supervised correlation coefficient obtained from random forest

Stephan Seifert · 2026

Read Understanding complex analytical data by a supervised correlation coefficient obtained from random forest by Stephan Seifert (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/PMC13585831.txt.

Introduction

Stephan Seifert explains a data method for reading complex chemical measurements in this 2026 paper. Instruments such as mass spectrometers produce hundreds of readings per sample, and sorting samples is only half the job: knowing which readings drive the sorting matters just as much. His focus is a method called Surrogate Minimal Depth, built on random forest, a model that grows many decision trees.

Random forests hide stand-in variables inside their trees, and this method uses those stand-ins to judge not only each reading's weight but also how readings sway one another. That joint sway is framed as a supervised correlation score, tuned by the sorting task itself. Worked examples on several kinds of analytical data show how the method picks variables and maps their links for a clearer read of messy datasets.

Catalog record

CREATORStephan Seifert
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE23589
LICENSEhttps://creativecommons.org/licenses/by/4.0
RIGHTSCC BY
SOURCEEurope PMC

Source: https://europepmc.org/article/MED/PMC13585831

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Understanding complex analytical data by a supervised correlation coefficient obtained from random forest Authors: Stephan Seifert Published: 2026 ABSTRACT The analysis of complex analytical data, such as that obtained from mass spectrometry, nuclear magnetic resonance, or vibrational spectroscopic approaches, necessitates the employment of multivariate chemometric methods for the classification of biological samples or the quantification of individual components. In this context, accurate prediction of the samples is not the sole aim. The interpretation of the variables that contribute to this prediction is also of utmost importance. Consequently, a range of algorithms, e.g., based on partial least squares regression or random forest, have been developed for variable selection. One of these approaches, Surrogate Minimal Depth (SMD), utilizes surrogate variables, which are obtained in random forest, to incorporate variable relationships into the analysis of variable importance and to examine the mutual impact of variables on the model. The latter is expressed as a parameter that can be regarded as a supervised correlation coefficient. This paper aims to lucidly illustrate the selection of variables and the analysis of variable relationships with SMD. Furthermore, the method’s potential for analyzing complex analytical data will be highlighted by demonstrating its application to various types of analytical data.Graphical abstract Introduction Complex biological materials, such as food, can be comprehensively characterized and classified using analytical techniques, which are, for example, based on mass spectrometry or various spectroscopic techniques. The data generated in this process is high-dimensional, meaning it is characterized by a large number of different variables from a relatively small number of samples. To exploit this data, multivariate methods are usually employed, meaning that all the variables in the dataset are analyzed simultaneously. These methods fall into two categories: unsupervised methods like principal component analysis, which are applied without additional information, and supervised methods, in which models are trained using samples with known assignments. In the latter case, it often makes sense to compare the performance of different algorithms, as their varying characteristics can affect the analysis differently depending on the specific features of the dataset being examined [1]. Supervised approaches based on latent variables, such as partial least squares regression (PLS), have the advantage of handling highly correlated data very well [2]. This is often advantageous when working with vibrational spectroscopic data. The non-parametric approach, random forest (RF), which is based on a large number of binary decision trees, also offers several distinct advantages, particularly when dealing with high-dimensional data [3]. RF has been demonstrated to be particularly effective in scenarios involving small sample sizes and a large number of uninformative variables [4]. Furthermore, RF provides internal validation. As each decision tree is built using a different subset of the training data generated via bootstrapping, the same dataset can be used for both training and validation simultaneously. For validation, only the “out-of-bag” samples, which were not used to train a given decision tree, are employed. Since random forests are much less tunable than other machine learning methods [5], meaning optimizing the hyperparameters has a relatively small impact on model performance, it is beneficial to omit this step. This has the huge advantage that the out-of-bag error from internal validation closely approximates the prediction error of an independent test dataset. Consequently, when analyzing a high-dimensional dataset, it is possible to use all samples for training while still obtaining a realistic estimate of the model’s performance on unseen data. Due to the advantages of RF, it is applied in many fields, including food authentication [6–8] and health data analytics [9, 10]. When applying supervised multivariate methods, it is not only important to obtain models with high predictive performance on unknown samples, but also to understand the underlying relationships that lead to these predictions. When PLS is used, the loadings can be used for interpretation. In addition, various methods have been developed for the selection of relevant variables [11]. Similarly, also, RF-based methods enable the differentiation between important and unimportant variables [12]. All of these methods determine the importance of each variable individually and do not take into account the relationships between them. The methodological expansion of Surrogate Minimal Depth (SMD) is that, rather than considering the variables individually, it analyzes their mutual impact on the RF model [13]. This concerns the assessment of the importance of variables in the random forest model, in which variable relationships are taken into account. Furthermore, the agreement of the variables in the individual splits of the decision trees is evaluated across the RF to calculate a relation parameter that quantifies the mutual impact of the features. The analysis of this parameter goes beyond a simple analysis of the importance of individual variables and enables a comprehensive interpretation of the relationships underlying the model. This, in turn, provides an overall characterization of the variable and sample properties involved. This paper provides a comprehensive explanation of SMD, focusing on diagrams rather than formulas. It also highlights various applications in which this analysis has revealed interesting relationships. Surrogate Minimal Depth (SMD) Surrogate variables were already included in Breiman’s original implementation of the random forest algorithm [3]. When training the decision trees that comprise a random forest, alternative splits were identified that could serve as optimal replacements for the original splits. This enables the model to be applied to datasets with missing variables (see Fig. 1a, left), which can be the case, for example, with medical patient data. In SMD, however, surrogate variables are used for a different purpose [13]. They are combined with the metric minimal depth, which determines the relevance of a variable based on the position at which it first appears in the decision trees (Fig. 1a, right). In SMD, the importance of a variable is thus determined by its initial appearance as a primary split or surrogate variable.Fig. 1Graphical explanation of the basic functioning of surrogate variables and SMD. a shows that a decision tree can be used to predict a dataset (S5) with missing values when surrogate variables are employed (left). Furthermore, a single decision tree is used as an example to show how the importance measures “Minimal Depth” and “Surrogate Minimal Depth” are determined (right). b illustrates the parameters “agreement” and “adjusted agreement,” which are used to determine surrogate variables and splits. While the agreement compares the performance of potential surrogates with that of the original split, the adjusted agreement also makes a comparison with the majority rule, an alternative approach of dealing with missing values in decision trees. c For relation analysis in SMD, the adjusted agreement for the respective combinations of variables (in this case, variables X4 and X2) is averaged across the entire RF, resulting in the mean adjusted agreement (MAA), which can be seen as a supervised correlation coefficient Graphical explanation of the basic functioning of surrogate variables and SMD. a shows that a decision tree can be used to predict a dataset (S5) with missing values when surrogate variables are employed (left). Furthermore, a single decision tree is used as an example to show how the importance measures “Minimal Depth” and “Surrogate Minimal Depth” are determined (right). b illustrates the parameters “agreement” and “adjusted agreement,” which are used to determine surrogate variables and splits. While the agreement compares the performance of potential surrogates with that

Read the complete text