app_bk.py

import marimo

__generated_with = "0.14.16"
app = marimo.App()


@app.cell
def _():
    import marimo as mo
    return (mo,)


@app.cell
def _(mo):
    mo.center(mo.md("# Home Credit Default Risk Prediction"))
    return


@app.cell
def _():
    import pandas as pd

    from sklearn.ensemble import RandomForestClassifier
    from sklearn.linear_model import LogisticRegression
    from sklearn.metrics import roc_auc_score
    from sklearn.model_selection import RandomizedSearchCV

    from sklearn.pipeline import Pipeline
    from sklearn.compose import ColumnTransformer
    from sklearn.impute import SimpleImputer
    from sklearn.preprocessing import MinMaxScaler, OneHotEncoder, OrdinalEncoder

    from lightgbm import LGBMClassifier

    from src.plots import (
        plot_target_distribution,
        plot_credit_amounts,
        plot_education_levels,
        plot_occupation,
        plot_family_status,
        plot_income_type,
    )
    from src.utils import get_dataset, get_features_target, get_train_test_sets
    from src.preprocessing import preprocess_data_pipeline
    return (
        get_dataset,
        get_features_target,
        get_train_test_sets,
        pd,
        plot_credit_amounts,
        plot_education_levels,
        plot_family_status,
        plot_income_type,
        plot_occupation,
        plot_target_distribution,
        preprocess_data_pipeline,
    )


@app.cell
def _(get_dataset, get_features_target):
    df = get_dataset()
    X, y = get_features_target(df)
    return X, df, y


@app.cell
def _(mo):
    mo.md("""## 1. Exploratory Data Analysis""")
    return


@app.cell
def _(mo):
    mo.callout(
        kind="info",
        value=mo.md(
            """💡 **Want a step-by-step walkthrough instead?**   
        Check the Jupyter notebook version here: 👉 [Jupyter notebook](https://huggingface.co/spaces/iBrokeTheCode/Home_Credit_Default_Risk_Prediction/blob/main/tutorial_app.ipynb)""",
        ),
    )
    return


@app.cell
def _(mo):
    mo.md("""### 1.1 Dataset Information""")
    return


@app.cell
def _(mo):
    mo.md("""**a. Shape of the train and test datasets**""")
    return


@app.cell
def _(X_test, X_train, df):
    train_samples = "Train dataset samples: {}".format(X_train.shape[0])
    test_samples = "Test dataset samples: {}".format(X_test.shape[0])
    columns_number = "Number of columns: {}".format(df.shape[1])

    train_samples, test_samples, columns_number
    return


@app.cell
def _(mo):
    mo.md("""**b. Dataset features**""")
    return


@app.cell
def _(X):
    X.columns
    return


@app.cell
def _(mo):
    mo.md("""**c. Sample from dataset**""")
    return


@app.cell
def _(X):
    sample = X.head(5).T
    sample.columns = [
        str(col) for col in sample.columns
    ]  # fix integer name warning
    sample = sample.astype(str)  # avoid numeric conversion issues in viewer
    sample
    return


@app.cell
def _(mo):
    mo.md("""**d. Target variable Distribution**""")
    return


@app.cell
def _(df, plot_target_distribution):
    target_table, target_plot = plot_target_distribution(df=df)
    target_table
    return (target_plot,)


@app.cell
def _(target_plot):
    target_plot
    return


@app.cell
def _(mo):
    mo.md("""**e. Number of columns of each data type**""")
    return


@app.cell
def _(X):
    X.dtypes.value_counts().sort_values(ascending=False)
    return


@app.cell
def _(X):
    categorical_cols = (
        X.select_dtypes(include=["object"]).nunique().sort_values(ascending=False)
    )
    categorical_cols
    return


@app.cell
def _(mo):
    mo.md("""**f. Missing data**""")
    return


@app.cell
def _(X, pd):
    missing_count = X.isna().sum().sort_values(ascending=False)
    missing_percentage = (missing_count / X.shape[0] * 100).round(2)

    missing_data = pd.DataFrame(
        data={"Count": missing_count, "percentage": missing_percentage}
    )
    missing_data
    return


@app.cell
def _(mo):
    mo.md("""### 1.2 Distribution of Variables""")
    return


@app.cell
def _(mo):
    mo.md(
        r"""Want to see how these plots were created? You can find the source code for the visualizations in [plots.py](https://huggingface.co/spaces/iBrokeTheCode/Home_Credit_Default_Risk_Prediction/blob/main/src/plots.py)."""
    )
    return


@app.cell
def _(mo):
    mo.md("""**a. Credit Amounts**""")
    return


@app.cell
def _(X, plot_credit_amounts):
    plot_credit_amounts(df=X)
    return


@app.cell
def _(mo):
    mo.md("""**b. Education Level of Credit Applicants**""")
    return


@app.cell
def _(X, plot_education_levels):
    education_table, education_plot = plot_education_levels(df=X)
    education_table
    return (education_plot,)


@app.cell
def _(education_plot):
    education_plot
    return


@app.cell
def _(mo):
    mo.md("""**c. Ocupation of Credit Applicants**""")
    return


@app.cell
def _(X, plot_occupation):
    occupation_table, occupation_plot = plot_occupation(df=X)
    occupation_table
    return (occupation_plot,)


@app.cell
def _(occupation_plot):
    occupation_plot
    return


@app.cell
def _(mo):
    mo.md("""**d. Family Status of Applicants**""")
    return


@app.cell
def _(X, plot_family_status):
    family_status_table, family_status_plot = plot_family_status(df=X)
    family_status_table
    return (family_status_plot,)


@app.cell
def _(family_status_plot):
    family_status_plot
    return


@app.cell
def _(mo):
    mo.md("""**e. Income Type of Applicants by Target Variable**""")
    return


@app.cell
def _(df, plot_income_type):
    plot_income_type(df=df)
    return


@app.cell
def _(mo):
    mo.md("""## 2. Preprocessing""")
    return


@app.cell
def _(mo):
    mo.md("""**a. Separate Train and Test Datasets**""")
    return


@app.cell
def _(X, get_train_test_sets, y):
    X_train, y_train, X_test, y_test = get_train_test_sets(X, y)
    X_train.shape, y_train.shape, X_test.shape, y_test.shape
    return X_test, X_train


@app.cell
def _(mo):
    mo.md("""**b. Preprocess Data**""")
    return


@app.cell
def _(mo):
    mo.md(
        r"""
    This preprocessing perform:

    - Correct outliers/anomalous values in numerical columns (`DAYS_EMPLOYED` column).
    - Encode string categorical features (`dtype object`).
        - If the feature has 2 categories, Binary Encoding is applied.
        - One Hot Encoding for more than 2 categories.
    - Impute values for all columns with missing data (using median as imputing value).
    - Feature scaling with Min-Max scaler

    Want to see how the dataset was processed? You can find the code for the preprocessing steps in [preprocessing.py](https://huggingface.co/spaces/iBrokeTheCode/Home_Credit_Default_Risk_Prediction/blob/main/src/preprocessing.py).
    """
    )
    return


@app.cell
def _(X_test, X_train, preprocess_data_pipeline):
    train_data, test_data = preprocess_data_pipeline(
        train_df=X_train, test_df=X_test
    )
    train_data.shape, test_data.shape
    return


@app.cell
def _(mo):
    mo.md("""## 3. Training Models""")
    return


@app.cell
def _(mo):
    mo.md(
        r"""At this points, we will work with `train_data` and `test_data` as features sets; also `y_train` and `y_test` as target sets."""
    )
    return


@app.cell
def _(mo):
    mo.md(r"""### 3.1 Logistic Regression""")
    return


@app.cell
def _(mo):
    mo.callout(
        mo.md("""
    In Logistic Regression, C is the inverse of regularization strength:

    - **Small C** → Stronger regularization → Simpler model, less overfitting risk, but may underfit.
    - **Large C** → Weaker regularization → Model fits training data more closely, but may overfit.
    """),
        kind="info",
    )
    return


@app.cell
def _(mo):
    mo.md(
        r"""
    We trained our Logistic Regression model using the following code:

    ```py
    # 📌 Logistic Regression
    log_reg = LogisticRegression(C=0.0001)
    log_reg.fit(train_data, y_train)

    # Train data predicton (class 1)
    lr_train_pred = log_reg.predict_proba(train_data)[:, 1]

    # Test data prediction (class 1)
    lr_test_pred = log_reg.predict_proba(test_data)[:, 1]

    # Get the ROC AUC Score on train and test datasets
    log_reg_scores = {
        "train_score": roc_auc_score(y_train, lr_train_pred),
        "test_score": roc_auc_score(y_test, lr_test_pred),
    }
    log_reg_scores
    ```

    📈 The ROC AUC scores obtained:
    """
    )
    return


@app.cell
def _():
    lr_scores = {
        "train_score": 0.6868418961663535,
        "test_score": 0.6854973003347028,
    }
    lr_scores
    return


@app.cell
def _(mo):
    mo.md(r"""### 3.2 Random Forest Classifier""")
    return


@app.cell
def _(mo):
    mo.md(
        r"""
    We trained our Random Forest Classifier model using the following code:

    ```py
    # 📌 Random Forest Classifier
    rf = RandomForestClassifier(random_state=42, n_jobs=-1)
    rf.fit(train_data, y_train)

    rf_train_pred = rf.predict_proba(train_data)[:, 1]
    rf_test_pred = rf.predict_proba(test_data)[:, 1]

    rf_scores = {
        "train_score": roc_auc_score(y_train, rf_train_pred),
        "test_score": roc_auc_score(y_test, rf_test_pred),
    }
    rf_scores
    ```

    📈 The ROC AUC scores obtained:
    """
    )
    return


@app.cell
def _():
    rf_scores = {"train_score": 1.0, "test_score": 0.7066811557903828}
    rf_scores
    return


@app.cell
def _(mo):
    mo.md(r"""### 3.3. Randomized Search with Cross Validations""")
    return


@app.cell
def _(mo):
    mo.md(
        r"""
    We trained the Randomized Search CV using the following code:

    ```py
    # 📌 RandomizedSearchCV
    param_dist = {"n_estimators": [50, 100, 150], "max_depth": [10, 20, 30]}

    rf_optimized = RandomForestClassifier(random_state=42, n_jobs=-1)
    rscv = RandomizedSearchCV(
        estimator=rf_optimized,
        param_distributions=param_dist,
        n_iter=5,
        scoring="roc_auc",
        cv=3,
        random_state=42,
        n_jobs=-1,
    )

    rscv.fit(train_data, y_train)

    rfo_train_pred = rscv.predict_proba(train_data)[:, 1]
    rfo_test_pred = rscv.predict_proba(test_data)[:, 1]

    rfo_scores = {
        "train_score": roc_auc_score(y_train, rfo_train_pred),
        "test_score": roc_auc_score(y_test, rfo_test_pred),
    }
    rfo_scores
    ```

    📈 The ROC AUC scores obtained:
    """
    )
    return


@app.cell
def _():
    rfo_scores = {
        "train_score": 0.8196620915431655,
        "test_score": 0.7308385425476998,
    }
    rfo_scores
    return


@app.cell
def _(mo):
    mo.md(r"""🥇The best results:""")
    return


@app.cell
def _():
    optimized_results = {
        "best_params_": {"n_estimators": 100, "max_depth": 10},
        "best_score_": 0.7296259755147781,
        "best_estimator_": "RandomForestClassifier(max_depth=10, n_jobs=-1, random_state=42)",
    }
    optimized_results
    return


@app.cell
def _(mo):
    mo.md(r"""### 3.4 LightGBM""")
    return


@app.cell
def _(mo):
    mo.md(
        r"""
    We trained our LightGBM Classifier model using the following code:

    ```py
    # 📌 LightGBM
    import warnings

    warnings.filterwarnings(
        "ignore", message="X does not have valid feature names"
    )

    # 📌 Get numerical and categorical variables (binary and mutiple)
    num_cols = X_train.select_dtypes(include="number").columns.to_list()
    cat_cols = X_train.select_dtypes(include="object").columns.to_list()

    binary_cols = [col for col in cat_cols if X_train[col].nunique() == 2]
    multi_cols = [col for col in cat_cols if X_train[col].nunique() > 2]

    # 📌 [1] Create the pipelines for different data types
    numerical_pipeline = Pipeline(
        steps=[
            ("imputer", SimpleImputer(strategy="median")),
            ("scaler", MinMaxScaler()),
        ]
    )

    binary_pipeline = Pipeline(
        steps=[
            ("imputer", SimpleImputer(strategy="most_frequent")),
            ("ordinal", OrdinalEncoder()),
            ("scaler", MinMaxScaler()),
        ]
    )

    multi_pipeline = Pipeline(
        steps=[
            ("imputer", SimpleImputer(strategy="most_frequent")),
            (
                "onehot",
                OneHotEncoder(handle_unknown="ignore", sparse_output=False),
            ),
            ("scaler", MinMaxScaler()),
        ]
    )

    # 📌 [2] Create the preprocessor using ColumnTransformer
    preprocessor = ColumnTransformer(
        transformers=[
            ("binary", binary_pipeline, binary_cols),
            ("multi", multi_pipeline, multi_cols),
            ("numerical", numerical_pipeline, num_cols),
        ],
        remainder="passthrough",
    )

    # 📌 [3] Create the Final Pipeline that combines the preprocessor and the model
    lgbm = LGBMClassifier(
        n_estimators=500,
        learning_rate=0.05,
        max_depth=-1,
        random_state=42,
        class_weight="balanced",
        n_jobs=-1,
    )

    lgbm_pipeline = Pipeline(
        steps=[("preprocessor", preprocessor), ("classifier", lgbm)]
    )

    # 📌 [4] Fit the Final Pipeline on the ORIGINAL, unprocessed data
    # The pipeline takes care of all the preprocessing internally.
    lgbm_pipeline.fit(X_train, y_train)

    lgbm_train_pred = lgbm_pipeline.predict_proba(X_train)[:, 1]
    lgbm_test_pred = lgbm_pipeline.predict_proba(X_test)[:, 1]

    lgbm_scores = {
        "train_score": roc_auc_score(y_train, lgbm_train_pred),
        "test_score": roc_auc_score(y_test, lgbm_test_pred),
    }
    lgbm_scores
    ```

    📈 The ROC AUC scores obtained:
    """
    )
    return


@app.cell
def _():
    lgbm_scores = {
        "train_score": 0.8523466410959462,
        "test_score": 0.7514895868142193,
    }
    lgbm_scores
    return


@app.cell
def _(mo):
    mo.md(r"""## 4. Model Performance Analysis""")
    return


@app.cell
def _(mo):
    lg_stat = mo.stat(
        label="Logistic Regression",
        bordered=True,
        value="🏋️ 0.687 🔎 0.685",
        caption="Scores are consistent across train and test, indicating no overfitting. However, the overall AUC is low, suggesting underfitting — the model is too simple to capture complex patterns.",
        direction="decrease",
    )

    rfc_stat = mo.stat(
        label="Random Forest Classifier",
        bordered=True,
        value="🏋️ 1.0 🔎 0.707",
        caption="Perfect training AUC indicates severe overfitting — the model memorized the training set. While the test score is better than Logistic Regression, the gap is too large for good generalization.",
        direction="decrease",
    )

    rfo_stat = mo.stat(
        label="Random Forest with Randomized Search",
        bordered=True,
        value="🏋️ 0.820 🔎 0.731",
        caption="Hyperparameter tuning greatly reduced overfitting. The smaller train–test gap and improved test AUC show better generalization and a strong performance.",
        direction="increase",
    )

    lgbm_stat = mo.stat(
        label="LightGBM",
        bordered=True,
        value="🏋️ 0.852 🔎 0.751",
        caption="Best overall performance. Small train–test gap and highest test AUC indicate a well-balanced model with strong generalization.",
        direction="increase",
    )

    mo.vstack(
        items=[
            mo.hstack(items=[lg_stat, rfc_stat], widths="equal", gap=1),
            mo.hstack(items=[rfo_stat, lgbm_stat], widths="equal", gap=1),
        ],
        gap=1,
        heights="equal",
        align="center",
        justify="center",
    )
    return


@app.cell
def _(mo):
    mo.md(r"""## 5. Model Selection""")
    return


@app.cell
def _(mo):
    mo.md(
        r"""
    Based on a comparison of all the models, the final model selection is clear.

    | Model | Train Score (AUC ROC) | Test Score (AUC ROC) |
    | :--- | :---: | :---: |
    | Logistic Regression | 0.687 | 0.685 |
    | Random Forest Classifier | 1.000 | 0.707 |
    | Randomized Search (Tuned RF) | 0.820 | 0.731 |
    | **LightGBM** | 0.852 | **0.751** |

    * The **Logistic Regression** model performed poorly due to underfitting.
    * The base **Random Forest** model, while better, suffered from severe overfitting.
    * The tuned **Random Forest** model was a significant improvement and a strong contender, achieving a solid `test_score`.
    * However, the **LightGBM** model ultimately demonstrated the best performance, achieving the highest **ROC AUC test score of 0.751**. This indicates that it is the most robust and accurate model for predicting loan repayment risk on unseen data.
    """
    )
    return


@app.cell
def _(mo):
    mo.callout(
        kind="success",
        value="🥇 Therefore, we will select the LightGBM model as our final choice for deployment.",
    )
    return


if __name__ == "__main__":
    app.run()