AmExpert CodeLab

Credit Card default risk

Objective:- To come up with an ML Model to determine whether companies or individuals would be able to return the money lent on time.

• Competition link : https://www.hackerearth.com/challenges/competitive/amexpert-code-lab/machine-learning/credit-card-default-risk-5-95cbc85f/

Approach

• First encoding the categorical columns ['gender','owns_car','owns_house']

  • binary encoded (0 or 1)

• column ['occupation_type'] contains categorical data (16 unique values)

  • encoded with value 0 to 16 based on position ,e.x :'Laborers' to 2 and 'Drivers' to 4

• columns ['migrant_worker','prev_defaults','default_in_last_6_month'] are already binary values (0,1)

• Missing values filled with mode and mean of that cols for categorical and continues columns respectively

  • ['no_of_children','total_family_members','migrant_worker'] ==> mode() for missing values
  • ['no_of_days_employed','yearly_debt_payments','credit_score'] ==> mean() for missing values for respective column

• handeling Imbalance dataset

  • dataset was highly imbalanced (ratio of 0/1 was very big)
  • used Imblearn's oversampling (SMOTE) to create a balance dataset
    • original shape ~(41k) , new size ~(83k)

• dataset of this stage is kept as (x_train,y_train) --> phase1

• Used Standardization to scale each column (expect target col) to bring mean to 0 and standard deviation to 1 for that column

  • used StandardScaler from sklearn.preprocessing

• dataset at this stage is kept as (x_train_scaled,y_train) ---> phase2

• Applied PCA technique to reduce dimensions from 16 to 12 .

  • used PCA from sklearn.decomposition
  • used a iterative while loop to see how many columns can be reduced to retain 95% of variance , using covariance matrix of generated by PCA

• dataset at this stage is kept as (x_train_scaled_pca,y_train) ---> phase3

• y_train remains the same as it contins only targets (dependent variable)

• uses the same instances used while train datset to preprocess the test data(test.csv)

  • instances like , StandardScaler , PCA()
  • created 3 different test dataframe
    • test_x , test_x_scaled, test_x_scaled_pca ( just like training phase)

Testing of different ML classifier training datasets:

• Used Kfold and cross validation score and tried running training dataset on 5 different classifier,given next to them are the avg. f1 score from cross_val_score()

  • normal raw dataset (phase1 dataset)

    • Logistic Regression : 0.9378
    • KNeighborsClassifier : 0.8301
    • LinearSVC : 0.7213
    • Random forest : 0.9874
    • XGBoost : 0.9807

  • scaled dataset (phase2 dataset)

    • Logistic Regression : 0.9688
    • KNeighborsClassifier : 0.9704
    • LinearSVC : 0.9692
    • Random forest : 0.9874
    • XGBoost : 0.9810

  • (scaled + pca) dataset (phase3 dataset)

    • Logistic Regression : 0.9689
    • KNeighborsClassifier : 0.9703
    • LinearSVC : 0.9692
    • Random forest : 0.9810
    • XGBoost : 0.9721

Score / Result from online HackerEarth Judge

• here dataset == test dataset(test.csv)

  • normal raw dataset (test_x)

    • Logistic Regression : 0.8085
    • KNeighborsClassifier : 0.4681
    • LinearSVC : 0.8085
    • Random forest : 0.91798
    • XGBoost : 0.8954

  • scaled dataset (test_x_scaled)

    • Logistic Regression : 0.8827
    • KNeighborsClassifier : 0.8710
    • LinearSVC : 0.8827
    • Random forest : 0.91731
    • XGBoost : 0.8962

  • (scaled + pca) dataset (test_x_scaled_pca)

    • Logistic Regression : 0.8835
    • KNeighborsClassifier : 0.8615
    • LinearSVC : 0.8835
    • Random forest : 0.89368
    • XGBoost : 0.8816

final standings on leaderboard (Rank 93) of the competition : link to leaderboard : https://www.hackerearth.com/challenges/competitive/amexpert-code-lab/leaderboard/credit-card-default-risk-5-95cbc85f/

Extra steps to improve accuracy :-

• more EDA.
• Hyperparameter Tuning