MODELS

EDA

Here is presented the Exploratory Data Analysis both for the dataset with "duplicates" and without, as a way of further investigating the reason about these seemingly duped rows.

CatBoost

RandomForest

XGBoost

GradientBoosting

Stacked (GB + XGB + RF + CB -> LR)

Importing the File

Steps needed to import and change into a proper table format the Json file from the Kaggle link: https://www.kaggle.com/datasets/rohit265/loan-approval-dataset

Setting up the Environment

It is important to note that Conda must be set up and configured in the Preferences of the KNIME Conda Integration.

The packages used in this workflow are:

Python: catboost, numpy, matplotlib, pandas, scikit-learn, seaborn

R: cairo, clValid, ggally, rserve, tidyverse

Partitioning of the Dataset

The dataset was initially split using stratified sampling based on the target variable, Risk_Flag, into a training set with 201,600 observations (80% of the total) and a test set with 50,400 observations (20% of the total). This approach will enable the models to be evaluated on data that simulates real-world unseen scenarios.

Preprocessing the Training Set

This component includes all the preprocessing steps applied to the training set and saves the necessary metrics to ensure the same transformations can be consistently applied to the test set.

Preprocessing the Test Set

This component ensures that all preprocessing modifications applied to the training set are mirrored on the test set, enabling accurate evaluation of model performance on unseen data.

Training of the Final Best Model - RandomForest

Important Notice

Some metanodes and components are marked with asterisks (*) to denote varying levels of computational intensity:

• * : Moderate computational intensity.

• ** : Substantial computational intensity.

• *** : Maximum computational intensity.

Nodes or components without asterisks have no special computational considerations.