Summary

The provided content outlines an exploratory data analysis (EDA) process for Home Credit's Credit Risk Model Stability dataset, including data preparation, merging, and the use of a LightGBM model for binary classification.

Abstract

The blog post details the steps taken to perform EDA on Home Credit's dataset for assessing the stability of their credit risk model. It begins with an overview of the dataset, which is sourced from diverse data points and available in both .csv and .parquet formats. The author then proceeds to demonstrate code snippets for importing necessary libraries, setting data types, handling missing values, and merging multiple tables based on 'case_id'. The EDA involves data cleaning, transformation, and feature engineering, ensuring that categorical data is properly encoded. A LightGBM model is trained and validated using the prepared dataset, with parameters optimized for binary classification. The author concludes by inviting readers to try an AI service recommended for its cost-effectiveness compared to ChatGPT Plus (GPT-4).

Opinions

The author emphasizes the importance of data type casting and handling missing values as part of the data preparation process.
The use of LightGBM for model training suggests a preference for gradient boosting decision trees in credit risk modeling.
The author believes that the shared EDA process and model parameters serve as a valuable reference point for further model refinement.
By recommending an AI service, the author implies that it offers similar capabilities to ChatGPT Plus (GPT-4) at a more affordable price.

Home Credit — Credit Risk Model Stability — EDA

In this blog post, we will conduct Exploratory Data Analysis (EDA) on the dataset related to Home Credit’s Credit Risk Model Stability.

As part of their credit risk assessment, Home Credit uses various data points to evaluate the creditworthiness of applicants. In this analysis, we will explore the dataset related to the stability of Home Credit’s credit risk model.

Dataset Overview

The dataset contains information about:

This dataset contains a large number of tables as a result of utilizing diverse data sources and the varying levels of data aggregation used while preparing the dataset. Note: All files listed below are found in both .csv and .parquet formats.

Data can be downloaded from here . To get an overview about data, please go through it.

Let’s dive into the code without delay.

import pandas as pd
import polars as pl
import numpy as np
import pandas as pd
import lightgbm as lgb
from sklearn.model_selection import train_test_split
from sklearn.metrics import roc_auc_score

# Correct function definition
def set_table_dtypes_pandas(df):
    # Iterate over the DataFrame's columns to identify and cast columns
    for col_name in df.columns:
        # If the last letter of the column name is 'P' or 'A', cast to float
        if col_name[-1] in ('P', 'A'):
            df[col_name] = df[col_name].astype(float)
    return df


def convert_strings(df: pd.DataFrame) -> pd.DataFrame:
    for col in df.columns:
        if df[col].dtype.name in ['object', 'string']:
            df[col] = df[col].astype("string").astype('category')
            current_categories = df[col].cat.categories
            new_categories = current_categories.to_list() + ["Unknown"]
            new_dtype = pd.CategoricalDtype(categories=new_categories, ordered=True)
            df[col] = df[col].astype(new_dtype)
    return df

from google.colab import drive
drive.mount('/content/drive')

!unzip "/content/drive/MyDrive/Dataset/credit/home-credit-credit-risk-model-stability.zip" -d "/content/tmp"

df_train_base = pd.read_csv("/content/tmp/csv_files/train/train_base.csv")

df_train_base.isnull().sum()

case_id          0
date_decision    0
MONTH            0
WEEK_NUM         0
target           0
dtype: int64

df_train_base.head(5)

df_train_base['date_decision'] = pd.to_datetime(df_train_base['date_decision'])
df_train_base['year'] = df_train_base['date_decision'].dt.year
df_train_base['month'] = df_train_base['date_decision'].dt.month

df_train_base.head(5)
df_train_base.describe()
df_train_appl_prev = pd.read_csv("/content/tmp/csv_files/train/train_applprev_1_0.csv")

df_test_base = pd.read_csv("/content/tmp/csv_files/test/test_base.csv",low_memory=False)
df_test_base = set_table_dtypes_pandas(df_test_base)
test_static_0_0 = pd.read_csv("/content/tmp/csv_files/test/test_static_0_0.csv",low_memory=False)
test_static_0_0 = set_table_dtypes_pandas(test_static_0_0)
test_static_cb = pd.read_csv("/content/tmp/csv_files/test/test_static_cb_0.csv",low_memory=False)
test_static_cb = set_table_dtypes_pandas(test_static_cb)
test_person_1 = pd.read_csv("/content/tmp/csv_files/test/test_person_1.csv",low_memory=False)
test_person_1 = set_table_dtypes_pandas(test_person_1)
test_credit_bureau_b_2 = pd.read_csv("/content/tmp/csv_files/test/test_credit_bureau_b_2.csv",low_memory=False)
test_credit_bureau_b_2 = set_table_dtypes_pandas(test_credit_bureau_b_2)

# Merging all the tables together on 'case_id'
data_submission = df_train_base.merge(
    test_static_0_0[['case_id']+selected_static_cols], how='left', on='case_id'
).merge(
    test_static_cb[['case_id']+selected_static_cb_cols], how='left', on='case_id'
).merge(
    test_person_1_agg, how='left', on='case_id'
).merge(
    test_person_1_feats_2, how='left', on='case_id'
).merge(
    test_credit_bureau_b_2_feats, how='left', on='case_id'
)

We can use LGB model to train

lgb_train = lgb.Dataset(X_train, label=y_train)
lgb_valid = lgb.Dataset(X_valid, label=y_valid, reference=lgb_train)

params = {
    "boosting_type": "gbdt",
    "objective": "binary",
    "metric": "auc",
    "max_depth": 3,
    "num_leaves": 31,
    "learning_rate": 0.05,
    "feature_fraction": 0.9,
    "bagging_fraction": 0.8,
    "bagging_freq": 5,
    "n_estimators": 1000,
    "verbose": -1,
}

gbm = lgb.train(
    params,
    lgb_train,
    valid_sets=lgb_valid,
    callbacks=[lgb.log_evaluation(50), lgb.early_stopping(5)]
)

This serves as a starting point for reference as I continue refining my final model. More details on the model will be shared soon.