Used the data set produced VMware to create a propensity to respond model. The dataset was initially an imbalanced dataset will 707 features and 50,006 rows. Therefore, the Synthetic Minority Over-Sampling Technique was used to balance the data. Pre-processing was required as the data contained a lot missing values. To address this issue, variables with more than 60% missing values were removed. And for the rest of the variables with less than 60% missing values, mean and mode imputation was done. This brought down the number of features from 707 to 653.
As there were 653 variables, it was necessary to perform feature selection to reduce the number of variables. Regularized L1 Logistic Regression (LASSO) was used to select important variables to work with. This method reduced the number of variables to 90. After the variables selection, Random Forest, XGBoost was performed on the data set to evaluate the suitable model. As the problem of imbalance was addressed by the SMOTE technique, XGBoost yielded 98.9987% Recall.
As there is a small chance for the data to be imbalanced, recall might not be a good measure to evaluate a model. Therefore, Macro Averaging Metric was used. From the Macro-Average F-score, the Random forest Yielded 0.87 whereas the XGboost yielded a F-score of 1. Therefore, it was concluded that XGBoost was the best model for this dataset. Hence a Propensity to respond model was created which can now be used for B2B Marketing.