6.4.1 Extracting the Savepoint Results by R Code:

You can extract specific results and attributes from the Savepoint using R code. The iMESc is not required here. To do this, use the following steps:

# Reading the Savepoint
savepoint <- readRDS("savepoint.rds")
savepoint$saved_data     # To access all saved Datalists
names(savepoint$saved_data)     # To access the names of the saved Datalists

# Accessing a specific Datalist named "datalist_name"
datalist_name <- savepoint$saved_data[['datalist_name']]
datalist_name

# Accessing specific attributes within the Datalist
attr(datalist_name, "factors")       # To access the Factor-Attribute
attr(datalist_name, "coords")        # To access the Coords-Attribute
attr(datalist_name, "base_shape")    # To access the Base-Shape-Attribute
attr(datalist_name, "layer_shape")   # To access the Layer-Shape-Attribute
attr(datalist_name, "extra_shape")   # To access the Extra-Shapes-Attribute

# To extract saved models
attr(datalist_name, "som")                 # To access all SOM models saved in the Datalist
attr(datalist_name, "som")[["model_name"]] # To access a saved SOM model named 'model_name'

# To access other models, replace "som" with the corresponding model name:
# 'kmeans' (k-Means), 'nb' (Naive-Bayes), 'svm' (Support-Machine Vector), 'knn'(k-nearest neighbor), 'rf' (Random Forest),
# 'sgboost' (stochastic gradient boosting), 'xyf' (supervised som).

Note: Ensure that you specify the correct path and filename of your save-point file in the readRDS function. Modify “datalist_name” and “model_name” in the R code to access specific Datalists and saved models, respectively.

7.1.1 Upload

• Name the Datalist: Use the text widget to provide a name for the Datalist.

• Numeric-Attribute: Upload a .csv or .xlsx file containing the numeric variables. This file is mandatory and should include observations as rows and variables as columns.The first row must contain the variable headings, and the first column should have observation labels. Columns containing characters (text or mixed numeric and non-numeric values) will be automatically transferred to the Factor-Attribute.

Create a Datalist - Upload

• Factor-Attribute: Upload a .csv or .xlsx file containing categorical variables. This file should have observations as rows and categorical variables as columns. The first row must contain variable headings, and the first column should have observation labels. If the Factor-Attribute is not uploaded, the observation IDs will be used automatically. This attribute is crucial for labeling, grouping, and visualizing results based on factor levels. It can be replaced at any time with a new one using the “Replace Factor-Attribute” button

• Coords-Attribute: Upload a .csv or .xlsx file containing geographical coordinates. This file is optional for creating a Datalist but required for generating maps. The first column should contain the observation labels, the second column Longitude values, and the third column Latitude values (both in decimal degrees). The first row must contain the coordinate headings.

• Base-Shape: Upload a single R file containing the polygon shape to be used for generating maps, such as an oceanic basin shape. This optional file can be created using the SHP toolbox available in the pre-processing tools, which allows converting shapefiles (.shp, .shx, and .dbf files) into a single R file.

• Layer-Shape: Upload a single R file containing an additional shape layer, such as a continent shape, to be used when generating maps. This optional file can also be created using the SHP toolbox available in the pre-processing tools.

7.1.2 Use Example Data

This option allows users to explore the example data available in iMESc . After clicking on “Create a Datalist,” select the “Use example data” radio-button to proceed with Datalist insertion. This action will insert two Datalists from Araçá Bay, located on the southeastern coast of Brazil:

• envi_araca: Contains 141 samples with 9 environmental variables.

• nema_araca: Contains 141 samples with 194 free-living marine nematode species (southeastern coast of Brazil).

Both Datalists comprise five attributes: Numeric, Factor, Coords-Attribute, Base-Shape, and Layer-Shape. Studies that explored these data area include Corte et al., 2017, Checon et al., 2018 and Vieira et al., 2021.

Create a Datalist - Use example data

7.2.1 Rename Datalist

Change the name of a selected Datalist to enhance organization and clarity.

7.2.2 Merge Datalists

Combine two or more Datalists by columns or rows. This action affects both Numeric and Factor-Attribute data. When merging by rows, it also impacts the Coords-Attribute. Please note that saved models in one of the Datalists are not transferred to the merged Datalist.

7.2.3 Exchange Factor/Variables

The “Exchange Factors/Variables” functionality in iMESc allows you to seamlessly convert or transfer data between numeric and factor formats. This powerful tool provides flexibility in handling your data and enables smooth transitions between different data types.

1. From Datalist Selector: Select the source Datalist from which you want to exchange data.

2. From Attribute Selector: Within the selected Datalist, choose between the Numeric or Factor Attribute that you wish to convert or transfer.

3. To Datalist Selector: Select the target Datalist where you want to transfer or convert the data.

4. To Attribute Selector: Within the target Datalist, specify whether you want to convert the data to Numeric or Factor format.

Now, let’s look at the conversion options based on the selected attributes:

7.2.3.1 From Numeric…

• To Numeric: This option allows you to copy or transfer the selected numeric variables from the source Datalist to the target Datalist while preserving their numeric format.

• To Factor: Convert the selected numeric variables to factors. You can use the “cut” option to categorize the variables into specified bins or levels. If the “cut” option is unchecked, each unique value will become a separate factor level. Additionally, you have the flexibility to edit the names and order of the factor levels.

Numeric to Factor Convertion

7.2.3.2 From Factor…

• To Factor: With this option, you can copy or transfer the selected factors from the source Datalist to the target Datalist, maintaining their original factor format.

• To Numeric: This conversion allows you to convert the selected factors to numeric data before copying or moving them. You have two types of conversions available:

  • Binary: For each factor level, a single binary column is created, where 1 indicates the class of that observation.

  • Integer: A single column is created, representing the numeric (integer) representation of the factor levels (values).

Factor to Numeric conversion

7.2.4 Replace Factor-Attribute

The “Replace Factor-Attribute” option allows users to update existing Factor-Attributes within a Datalist by replacing them with new data from a CSV file.

7.2.5 Edit Datalist Columns

The “Edit Datalist Columns” feature enables users to modify the names of columns, including both Numeric-Attributes and Factor-Attributes, within a Datalist.

7.2.6 Edit/Merge models

This functionality allows you to:

• Edit model names: Select the target Datalist, the model type (e.g., random forest), enter a new name, and confirm the changes.

• Merge models: Select the models and the target Datalist. The selected models will be copied to the target Datalist, preserving the original models. In case of duplicate model names, iMESc automatically differentiates them by adding ‘.1’, ‘.2’, and so on.

7.2.7 Build IDs/Columns & Formula

Creates columns or build IDs in the Numeric or Factor-Attribute using three options:

• “Find & Replace”: The user selects a target variable/factor, specifies a pattern to find, and another pattern to replace. Then, the user can create a new column using the results. This option employs the gsub function from R base to find the specified pattern and replace it with the desired one throughout the observations.

  

• “Concatenate”: The user selects multiple target variables/factors, and like the Concatenate function in Excel, iMESc merges them. The user can build a new column or use the concatenated results as IDs for observations. This functionality in iMESc uses the paste0 function from R base.

• “Apply formula”: Allows the user to apply custom formulas. The user selects multiple target variables to be available in the bucket and builds a formula by dragging the targets and formula elements (e.g., (, ), +, -,*, etc.). The user can then create a new column with the formula results.

• 

7.2.8 Transpose a Datalist

This tool allows the rotation of a Datalist (Numeric and Factor) from rows to columns.

7.2.9 SHP toolbox

This toolbox allows the creation of Base-Shapes, Layer-Shapes and Extra shapes.

1. Upload shape files* at once
2. Select the shape attribution: Base-Shape, Layer-Shape or Extrashape
3. Select the Target Datalist
4. Include the shape or download a single R file to be uploaded later when creating a Datalist.

7.2.10 Datalist Manager

Manage saved Datalists and their attributes. The manager displays the size of each Datalist, and options for deleting attributes and downloading dataframes.

7.2.11 Delete Datalists

Remove a Datalist entirely.

8.2.1 Summary

The Summary tab consists of three buttons:

• Datalist: Shows the structure of the Datalist.

• Data: Displays the number of rows, columns, and missing values in the Numeric-Attribute.

• Variable: Provides summary statistics and histograms for each variable in the Numeric-Attribute.

• Factors: Presents a summary of the Factor-Attribute and a stacked plot with the count of Factor levels.

Summary

8.2.2 Ridges

This functionality plots the Numeric-Attribute as ridges lines grouped by a Factor.

Ridge plots are partially overlapping line plots that create the impression of a mountain range. They are useful for visualizing changes in distributions over a gradient. This function uses the geom_ridgeline function from the ggridges package.

Ridges

8.2.3 Pairplot

The pair plot feature in iMESc utilizes the ggpairs function from the GGally package to create a matrix of plots based on selected variables from the Numeric-Attribute, and optionally one additional categorical variable from the Factor-Attribute to group the ridges.

In a pair plot, each variable is plotted against all other variables in the dataset, resulting in a matrix of scatterplots showing the relationship between each pair of variables.

Users have the flexibility to customize the display in each panel according to their preferences:

• Upper Panel: Users can opt to display correlations or correlations per group or none.

• Lower Panel: Options include displaying points, points colored per group, or no points at all.

• Diagonal Panel: Users can choose from several options, such as Density, Density per group, Histogram, Histogram per group, or none.

8.2.4 Boxplot

• The Box Plot tab allows users to create boxplots with numerous graphical options.

MDS

8.2.5 Correlation plot

• The Correlation Plot tab uses the heatmap.2 function from gplots package. Users can choose the correlation method (e.g., Pearson, Spearman) and filter correlations based on a cutoff value (filtering done using the findCorrelation function from caret). All arguments available in this function are accessible in iMESc. For a detailed description of all arguments, users can refer to the documentation of the heatmap.2 function.
  

8.2.6 MDS

8.2.7 PCA

• The PCA tab utilizes the prcomp function from R base. We recommend scaling the data before conducting PCA analysis. Users can interact with several graphical options, including coloring observations by a factor from the associated Factor-Attribute.

PCA

8.2.8 RDA

The RDA tab employs the rda function from the vegan package. It requires at least two Datalists to build a Y~X model. The results are presented in two panels:

• Plot: Offers several graphical options, including coloring observations by a factor from the associated Factor-Attribute.

• Summary: Provides results of the rda model, such as variable importance (constrained and unconstrained), variable and observation scores, linear constraints, and biplot.

RDA

8.2.9 segRDA

The segRDA tab implements the segRDA workflow (Vieira et al., 2019) for modeling non-continuous linear responses of ecological data. The routine consists of three panels:

• SMW: Split Moving Windows along ordered data.

• DP: Dissimilarity Profile of the SMW results for identifying breakpoints.

• pwRDA: Performs a piece-wise redundancy analysis using specified breakpoints. The results are presented in the same way as in the RDA panel.

segRDA routine

8.4.1 Biological Diversity Indexes

8.4.2 Niche Analysis

The Niche Analysis is performed using the niche function from the ade4 package (Dray and Dufour, 2007). It involves running a Principal Component Analysis (PCA) on the environmental table and associating the resulting table of row profiles with the corresponding faunistic table. This association provides the average position (i.e., niche position) of each species along the ordination axes, allowing researchers to understand the ecological niche of each species in relation to the environmental factors.

The analysis also includes the Optimal Niche Position (OMI) analysis, which provides additional insights into the species’ niches. It calculates the niche breadth value, representing the total variance of the environmental table weighted by the species’ abundances. This information helps researchers understand the range of conditions under which each species thrives.

Moreover, iMESc enables the extraction of the niche position (NP), environmental boundaries (EBs), and niche breadths (NBs) of species along the principal components (Vieira et al., 2019). This detailed analysis allows for a deeper understanding of how species interact with the environment and their ecological preferences.

For more in-depth information about the calculation of niche parameters, researchers can refer to the work of Dolédec et al. (2000). The Niche Analysis is a valuable tool for ecologists and environmental scientists to study species-environment relationships and gain valuable insights into species distributions and their ecological roles within a given ecosystem.

8.5.1 Self-Organizing maps

The Self-Organizing Maps (SOM) module in iMESc offers users powerful functionality to perform Self-Organizing Maps analysis on their target Datalists. This analysis is implemented using the supersom function from the kohonen R package. By default, the method trains a single layer, but users can train multiple layers by activating the “supersom” button and fine-tune the distance for each layer, as well as adjust the weights for each layer. The SOM module is conveniently organized into three panels: Training, Results, and Predict, providing an intuitive and efficient workflow for analyzing and visualizing data.

8.5.2 Hierarchical clustering

Hierarchical Clustering (HC) analysis is performed on the Numeric-Attribute or SOM-codebook (if any saved SOM model). The process starts by assigning each object to its own cluster and then iteratively joins the two most similar clusters until there is only one cluster left. At each stage, the distances between clusters are recomputed using the Lance–Williams dissimilarity update formula based on the specific clustering method chosen.

iMESc utilizes the hcut function, part of the factoextra package, to implement HC analysis. The HC analysis can be performed on the Numeric-Attribute or on the codebook obtained from a saved SOM model. Users have the option to choose the specific clustering method that suits their analysis needs.

The HC analysis results can be saved using the flashing blue button in tabs 3 and 4, enabling users to conveniently capture and analyze their clustering results for further exploration or other analytical tasks in iMESc. This feature enhances the usability of iMESc and provides users with a seamless workflow for conducting hierarchical clustering and utilizing the clustering outcomes in various downstream analyses.

Users can customize several parameters in iMESc:

8.5.3 K-means

The K-means module in iMESc performs clustering on Numeric-Attribute or SOM-Codebook data from a selected Datalist. The objective of K-means clustering is to divide data points into K clusters while optimizing the within-cluster sum of squares. iMESC offers three K-means algorithms: “Hartigan-Wong,” “Lloyd-Forgy,” and “MacQueen.” Each algorithm has distinct initialization and classification methods. Details about the algorithms can be found here.

8.6.1 Key features of iMESc’s Supervised Techniques

In Supervised module of iMESc, researchers have access to a diverse range of machine learning algorithms for predictive modeling. Below is an overview of the available algorithms along with their respective model tags (used by the train function in caret), a brief description and the required packages utilized by iMESc via caret:

8.6.2 Naive Bayes

Definition: Naive Bayes is a probabilistic algorithm used for classification tasks in machine learning. It calculates the probability of each possible classification given the data, using Bayes’ theorem. The “naive” assumption is that all features (variables) in the data are independent of each other, which simplifies the calculations. Despite this assumption, Naive Bayes often performs well on a wide range of classification problems.

Algorithm: in iMESc, Naive Bayes is implemented using the ‘nb’ algorith from klarR package (via caret).

Advantages: Naive Bayes is computationally efficient and relatively simple to implement. It performs well with a small amount of data and can handle high-dimensional feature spaces. It has been widely applied in various areas, including environmental science, where it can be used for tasks like species classification, pollutant detection, and environmental monitoring.

Limitations: The “naive” assumption of independence between features can be limiting in certain cases where features are correlated. In such scenarios, other algorithms that consider feature dependencies may perform better.

Model hyperparameters implemented in iMESc:

References: Kalcheva et al., 2020

8.6.3 Support Machine Vector

Definition: Support Vector Machine (SVM) is a versatile algorithm used for both classification and regression tasks in machine learning. SVM aims to find the optimal hyperplane that separates classes in the data or predicts numeric values with maximum margin. It works by transforming the data into a higher-dimensional space and finding the best separating hyperplane that maximizes the margin between classes.

Algorithm: In iMESc, SVM is implemented using the ‘svmLinear’ and ‘svmRadial’ algorithms from the kernlab package (via caret).

Advantages: SVM is effective in high-dimensional spaces, making it suitable for datasets with many features. It can handle both linear and non-linear relationships between features and the target variable. SVM is less prone to overfitting, especially with the ‘svmRadial’ kernel, which can capture complex patterns in the data (Cervantes et al., 2020).

Limitations: SVM can be computationally intensive, especially with large datasets. It may also be sensitive to the choice of kernel and hyperparameters, requiring careful tuning for optimal performance (Cervantes et al., 2020).

Model hyperparameters implemented in iMESc:

8.6.4 K-Nearest neighbor

Definition: K-Nearest Neighbors (KNN) is a non-parametric algorithm used for both classification and regression tasks in machine learning. It operates on the principle of similarity, where it assigns a new data point to the class or predicts its value based on the majority class or average value of its k-nearest neighbors in the feature space. KNN is intuitive and easy to understand, making it a popular choice for various applications.

Algorithm: In iMESc, KNN is implemented using the ‘knn’ algorithm from caret.

Advantages: KNN does not make assumptions about the underlying data distribution, making it useful for complex and non-linear relationships between features and the target variable. It can handle multi-class classification tasks and works well with both numerical and categorical data. KNN is relatively simple to implement and serves as a baseline algorithm for comparison with other complex models (Guo et al., 2003).

Limitations: KNN can be computationally expensive, especially with large datasets, as it requires calculating distances between the new data point and all training samples. The choice of the value for ‘k’ (the number of neighbors) significantly impacts the model’s performance. A small ‘k’ may lead to overfitting, while a large ‘k’ may lead to underfitting (Guo et al., 2003).

Model hyperparameters implemented in iMESc:

8.6.5 Random Forest

Definition: Random Forest is an ensemble learning algorithm used for both classification and regression tasks in machine learning. It operates by constructing multiple decision trees during training and combines their predictions to make more accurate and robust predictions. Each decision tree is constructed using a random subset of the training data and a random subset of features, introducing diversity among the trees.

Algorithm: In iMESc, Random Forest is implemented using the ‘rf’ algorithm from the RandomForest package (via caret).

Advantages: Random Forest is known for its high predictive accuracy, generalization ability, and resistance to overfitting. It can handle both numerical and categorical features without the need for extensive data preprocessing. Random Forest is robust to noisy data and can provide estimates of feature importance, helping researchers understand the relative importance of different features in the prediction (Gupta et al., 2017).

Limitations: One of the primary challenges users face is the complexity of the outcomes it produces, which can often make the interpretations non-intuitive, both in classification and regression tasks. Furthermore, it has a tendency to overfit when engaged with datasets containing noisy elements in classification or regression tasks, potentially affecting the precision of predictions on new data (Gupta et al., 2017).

Model hyperparameters implemented in iMESc:

8.6.6 Stochastic Gradient Boosting

Definition: Stochastic Gradient Boosting, also known as Gradient Boosting Machine (GBM), is an ensemble learning algorithm used for both regression and classification tasks in machine learning. GBM builds multiple weak learners (typically decision trees) sequentially, and each subsequent learner corrects the errors made by the previous one. This iterative process leads to improved model performance and generally increase predictive accuracy.

Algorithm: In iMESc, Stochastic Gradient Boosting is implemented using the ‘gbm’ algorithm from gbm package (via caret).

Advantages: GBM combines the strengths of both boosting and bagging techniques, offering a streamlined approach to data modeling. It alleviates the need for pre-selecting or transforming predictor variables, thereby simplifying the initial stages of model setup. Notably, it enhances prediction accuracy and computational speed by utilizing only a subset of the training data, thus avoiding overfitting while being resistant to outliers. The methodology, characterized by a sequential forward stagewise procedure, focuses on minimizing the loss function effectively, continuously improving the model’s accuracy through adaptive weighted resampling strategies (Shin, 2015; Yao et al., 2019).

Limitations: The process, although optimized for accuracy, tends to create complex models, making interpretation challenging and potentially increasing computational time as the model enlarges. Additionally, the constant adjustment of sampling weights at each iteration can sometimes escalate the model’s sensitivity to noisy data and outliers, which might undermine the model’s stability over time. Furthermore, the intense focus on misclassified observations during training might sometimes compromise its ability to generalize well to new, unseen data (Shin, 2015; Yao et al., 2019).

Model hyperparameters implemented in iMESc:

8.6.7 Supervised Self-Organizing Maps

Definition: Supervised Self-Organizing Maps (XYF) is a variant of the traditional Self-Organizing Maps adapted for supervised learning tasks. SOM is an unsupervised learning algorithm used for clustering, dimensionality reduction, and visualization of high-dimensional data. However, in the supervised version, additional information about the class labels or categories of the data is used to guide the map’s training.

Algorithm: In iMESc, Supervised Self-Organizing Maps are implemented using the ‘xyf’ algorithm from Kohonen package (via caret).

Advantages: The XYF Network excels in condensing multidimensional data into a user-friendly two-dimensional grid, facilitating the easy visualization and analysis of complex datasets. This network is adept at handling non-linear relationships between input and output spaces, circumventing common scaling issues seen in other models. Furthermore, its ‘open’ model structure allows users to visually explore and understand the underlying relationships in the data, fostering a more intuitive analysis process (Melssen et al., 2006).

Limitations: Users need to initially establish a balanced influence between the input and output objects, a process that can complicate the setup phase. Proper scaling of these variables is vital, and any missteps can lead to performance issues in the model, sometimes resulting in inaccurate predictions (Melssen et al., 2006).

Model hyperparameters implemented in iMESc:

8.6.8 Explore Saved Models

The “Explore Saved Models” module in iMESc allows users to explore and analyze various machine learning models, including NB, SVM, KNN, RF, GBM and XYF models. The module provides an interface for selecting and evaluating saved models on different datasets and making predictions on new data. The user can choose from the following prediction options:

1. Training Data: Predictions will be made on the training data on which the model was originally trained.

2. Partition (if any): Predictions will be made on the partitioned “Test” data that was set aside during the training phase for external validation.

3. Training + Partition (if any): Predictions will be made on both the training data and the partitioned “Test” data.

4. New Data: The user can select a new dataset (Datalist) on which predictions will be made.

The results are presented in six tabs:

1. Summary: This tab displays the print and plot default output of the selected model, including performance metrics along the tuning process, such as Accuracy and R-squared.
2. Performance: In this tab, performance metrics and the confusion matrix (for classification models) are shown to evaluate the model’s predictive performance.
3. Predictions: Sheet with the prediction for each observation. User can Create a Datalist with this result.
4. Permutation Importance: In this tab, users can perform a permutation importance analysis to assess the significance of predictors in the model. More details about this analysis are provided in the Permutation Importance section.
5. Partial Dependence: This tab generates a partial dependence biplot using the “partial” function from the ‘pdp’ package. The biplot helps visualize the partial dependence of the model on specific features.
6. Explore Model Features: In this tab, users can explore the importance of model features through various graphical outputs. For all algorithms, a pair plot is available. Additionally, specific visualizations are provided for each algorithm:

   • NB: Density plots for the predicted classes.

   • SVM: Feature importance based on SVM coefficients and feature values.

   • RF: Individual decision tree plot and Feature Importance based on Mean Decrease Gini.

   • GBM: Default feature plot model and Error vs. number of trees plot.

   • XYF: Codebook plot.

7. Resampling plot. This tab presents a bar plot showing how each observation was classified using the resampling predictions.

8.7.1 Compare

This submenu is designed for model comparison. In this tab, the user selects the Datalist containing the models (with similar resampling schemes) to be compared and explores the results through 8 tabs:

• 1.1. Summary: table with performance metrics.

• 1.2. bwplot: boxplot of performance metrics.

• 1.3. densityplot: creates density plots that show the distribution of each feature in the dataset, grouped by the levels of the target variable.

• 1.4. dotplot: creates a dot plot of the model performance metrics.

• 1.5. parallelplot: creates a plot where each axis represents a different performance metric, and the lines connecting the points for each model indicate its performance on each metric.

• 1.6. splom: creates scatterplot matrices, where each scatterplot in the matrix represents the relationship between two performance metrics, and the points for each model are plotted in the corresponding location in each scatterplot.

• 1.7. P-values (pairwise comparisons): computes pairwise comparisons between models using the t-test and returns a matrix containing the p-values for each comparison.

8.7.2 Ensemble

The “Ensemble” submenu in iMESc allows users to create ensemble models using the following steps:

1. Select the Datalist:

   • The user chooses a Datalist containing the saved models.

2. Select the Models:

   • The models saved in the selected Datalist will be grouped into pools based on their problem type (classification or regression) and Y (variables used as a response).

   • The user can individually select models from these lists to form an Initial Pool of models for the ensemble.

3. Prediction Type:

   • The user can choose from the following options:

     • New predictions + validation: This option generates ensemble predictions along with validation measures.

     • New predictions (unknown data): Make predictions on new, unknown data.

     • Use saved model: Utilize a previously saved ensemble model to make predictions.

4. Data for Predictions:

   • The user can choose one of the following options:

     • Partition (if any): Select a data partition used during model training for predictions.

     • New Data: Generate predictions for a user-selected Datalist containing new data.

5. Validation Data (if applicable):

   • This option is available when “New predictions + validation” is selected in step 3, and New Data is selected in step 4.

   • The user should select the Datalist containing the observed values to be compared with predicted values during the validation process.

6. Ensemble Method:

   • For classification, the user can choose from the following methods to combine individual models in the ensemble:

     • Majority Votes: Each individual model votes for a class, and the majority class wins.

     • Majority weighted votes (overall performance): Each classifier is assigned a weight based on its classification performance on a validation set, such as accuracy. The final decision is made by summing all weighted votes and selecting the class with the highest aggregate. The weights can be estimated using either the training data (i.e., the original X and Y data used to train the models) or a new data X and Y.

     • Majority weighted votes (adaptive): Proposed by Dogan & Birant (2019), this method gives more weight to models that correctly classify observations not correctly predicted by most classifiers. Weights can be estimated using the training data or a new data X and Y.

7. Estimate Weights Using:

   • This step determines how the weights calculation should be based.

   • If “New predictions + validation” is selected in the previous step, the user can choose between:

     • New data/validation set: Calculate weights based on new data or validation set.

     • Training Data: Calculate weights using the original training data.

8. Search Best Ensemble (if applicable):

   • This switch button is available when “New predictions + validation” is selected in step 3.

   • When enabled, iMESc generates predictions using combinations of models in the pool to search for the best combination that produces the highest performance score.

   • The user can choose between two functions to calculate metrics during the search process:

     • postResample: Calculate accuracy and Kappa using caret’s postResample function.

     • multiClassSummary: Calculate accuracy, Kappa, and various statistical averages (F1, Sensitivity, Specificity, Precision) using caret’s multiClassSummary function. Additional options with multiclass AUC (for classification) are also available (with computational cost).

9. Make Predictions:

   • A flashing button to run predictions based on the selected ensemble model settings.

   • This step executes the ensemble model creation process and generates predictions based on the chosen ensemble method and validation settings.

The results generated by iMESc are organized into nine tabs, each presenting different aspects of the ensemble model and its predictions. Here’s a description of each tab: