zerocool909 / analytics-vidhya-job-hackathon-may

Approach and code set for the competition

Analytics Vidhya hackathon May

https://datahack.analyticsvidhya.com/contest/job-a-thon-2/?utm_source=datahack&utm_medium=Navbar&utm_campaign=Jobathon

Approach and code set for the competition

JOB-A-THON - May 2021

Credit Card Lead Prediction

Analysis and Solution by Aditya Raj

Problem Statement:

Happy Customer Bank is a mid-sized private bank that deals in all kinds of banking products, like Savings accounts, Current accounts, investment products, credit products, among other offerings.

The bank also cross-sells products to its existing customers and to do so they use different kinds of communication like tele-calling, e-mails, recommendations on net banking, mobile banking, etc.

In this case, the Happy Customer Bank wants to cross sell its credit cards to its existing customers. The bank has identified a set of customers that are eligible for taking these credit cards.

Now, the bank is looking for your help in identifying customers that could show higher intent towards a recommended credit card, given: • Customer details (gender, age, region etc.) • Details of his/her relationship with the bank (Channel_Code,Vintage, 'Avg_Asset_Value etc.) Solution: The current solution is based on assumptions on EDA(graphs and other details included in Jupyter notebook .ipynb file) and ascertaining the data type as well as feature engineering.

1. Data collection : All data used for training is provided by the competition admins, no external data is used for training the model.
2. Data Pre-processing : A total of 10 features are provided. EDA study was carried out based on pandas_profiling on training set. Attached is av_credit_card_lead_eda.html for reference.

Categorical columns 0 ID 1 Gender 3 Region_Code 4 Occupation 5 Channel_Code 7 Credit_Product 9 Is_Active Numerical Columns 2 Age 6 Vintage 8 Avg_Account_Balance

Creating a basic heat map of the training set we find that though other features are nearly correlated with ‘Is_Lead’ but their removal depends on how they fit while modelling the data.

When observing all the the categorical columns, some data was found to be NaN for Credit_Product

We fill Credit_Product with ‘Unknown’ as there are missing values in test set too on which we need to predict.

Now when looking at numerical columns we observe that data corresponding to Vintage, Age and Avg_Account_Balance is skewed. Also, Avg_Account_Balance has few outliers,

• I tried both approaches of removing(values >0.999 and values<.001 quantile)as well as keeping these outliers. Though keeping them provided better AUC score.

For age, I tried binning it into 4 categories of young, middleage, aged, old as age_category. We observed the ls_Lead (0/1) value distribution was differing. Hence, the model can learn the distribution in bin category and provide accurate results.

I also tried binning Vintage also in years feature but that did not help the model accuracy.

All the categorical values were then label encoded into numerical values.

I also, tried the approach of one-hot encoding the categorical values but that did not improve the model. Also, Tree-models struggle if there are a large number of levels/features, regardless of how much data we have. Hence, keeping less features and label-encoded values.

Since, we had few skewed features in dataset we scale the training set and test set using StandardScalar.

As, this is a class imbalance problem, the data distribution is uneven between ‘Is_Lead’: (0/1) Approach used:

1. Oversampling :Impute synthetic data via SMOTE, so make Is_Lead-0 and Is_Lead :1 value to equal up. More synthetic data points for 1 was added.

2. Undersampling: Train with random selection of equal number of Is_Lead:0 to Is_Lead :1. So reduce higher volume of Is_Lead :0

3. Continue with maximum training so that model understands the Is_Lead:1 data points with more confidence The step 1 & 2 resulted in lower AUC score compared to step 3. Hence final model was trained without synthetic of equal valued for Is_Lead:1 and Is_Lead:0

4. Model Training : For model training, I have tried various classifiers such as Decision Tree, Random Forest, k-means, Logistic Regression, SVC , XGBoost, SGDClassifier with GridSeacrch and LightGBM with GridSearch. I have not tried ANN due to past experience of simple interpretable models performing better on such tasks. Among tried models, LightGBM gave best AUC Score and no overfitting. [.0.872743928714212] • I tried to train with maximum examples as possible to avoid overfitting and underfitting. Hence, distribution set of .9 for training and .1 for test was used. I did try out with distribution set of.7 and .3 but observed the AUC score get better with more training on provided data. Since the created model has same AUC score as on public dataset. I don’t think this approach has overfitted on data yet. • The model I have used is LightGbm with hyperparameter tuning using GridSearchCV for 4 cross validations on the dataset. • I tried ensemble modelling by majority voting ensemble using different hyperparameter value models for LightGBM. But, I am assuming the properly tuned model should be performing better on private set. • I have not tried other ensemble combination for LightGBM with other models like Random forest, XGBoost etc, Model results: Overall accuracy of Light GBM model: 0.8609449395678184

• I have also tried with few top features as feature_fraction=0.7 in model params

AUC score: 0.8695915099762787

4.Solution submission: The best performing model is used to predict_proba for probability values of both classes for ‘Is_Lead’. The vales corresponding to 1 is used for prediction.