Diviner

Page under construction

Diviner

Diviner is a semi-automated machine learning (Semi-AutoML) tool specifically designed to enhance the development of multivariate calibration models for linear regression. Unlike traditional AutoML systems that entirely automate the machine learning workflow, often at the expense of domain-specific insights and transparency, Diviner strikes a balance between automation and expert involvement. It allows users to leverage automation efficiently while maintaining control over critical decision points in the modeling process. This hybrid approach addresses key shortcomings of AutoML, such as the lack of domain knowledge, overfitting, and limited customization, by integrating user input to guide model development more effectively.

Diviner is a tool designed for calibrating linear models, specifically Partial Least Squares (PLS) and regularized multiple linear regression (MLR) models, such as Elastic Net. It provides a comprehensive workflow that includes outlier assessment, a grid search for preprocessing methods, variable selection, and user-guided model refinement.

Due to its extensive library of preloaded preprocessing methods, Diviner is particularly suited for chemometrics and spectral data analysis. However, users can also create custom preprocessing libraries, making Diviner a versatile tool for linear regression tasks with various data types.

Data Workflow

Exploratory Module:

Data loading: The process begins with data being loaded into the system. If a test dataset is available, it should be loaded at this time so it can be used to evaluate the performance of the models in that test dataset. Note: if a test set is not loading before the run, applying the models to the test set a posteriori won't be possible.

Choice of algorithms: PLS is set by default, and MLR must be activated in the options. While the optimization of the PLS models is part of the initial grid search in diviner, MLR (Elastic Nets) on the other hand, performs an optimization routine to find the best penalty for each model.

Preprocessing: This step involves selecting multiple preprocessing methods depending on the data type and application. In the same step, preprocessing methods used for outlier assessment must be selected.

Cross-validation (CV): Diviner supports all modes of cross-validation in the PLS_Toolbox and Solo. This step also sets the number of PLS latent variables (LVs) that will be used in the initial grid search.

Auto-Variable Selection: Variable selection in the first module uses two fast algorithms VIP (Variable Importance in Projection) and sRatio (Selectivity Ratio). https://www.wiki.eigenvector.com/index.php?title=Selectvars

Outlier assessment: Diviner automatically assesses possible outliers in the dataset using a combination of robust PCA and PLS with different preprocessing methods previously chosen by the user.

Preliminary Output & First Model Selection: after the full grid-search of PLS (and MLR) models based on preprocessing recipes, variable selection, and Latent Variable combinations. The initial output is generated as a plot of the Overfit (RMSECV/RMSEC) vs. RMSECV. The user manually selects the best models from this plot for refinement or further analysis.

Refinement Module:

Refinement consists of further variable selection and outlier inclusion. Since these two procedures are time consuming and not ideal to perform on a large number of models it follows the model selection from the grid-search on the first module.

Further Variable Selection: In this step, additional refinement of the variable selection is carried out, using interval Partial Least Squares (iPLS) to fine-tune which variables contribute to the model.

Outlier Reinclusion: If any outliers previously excluded are significant upon further analysis, they might be re-evaluated and potentially reintegrated into the models.

Final Model Selection and Output: After all refinements are made, the best-performing models are selected by the user for the final output.

Text

File:Image file name.png