A Comparison of Missing Value Imputation Techniques on Coupon Acceptance Prediction

Abstract

The In-Vehicle Coupon Recommendation System is a type of coupon used to represent an idea of different driving scenarios to users. Basically, with the help of presenting the scenarios, the people’s opinion is taken on whether they will accept the coupon or not. The coupons offered in the survey were for Bar, Coffee Shop, Restaurants, and Take Away. The dataset consists of various attributes that capture precise information about the clients to give a coupon recommendation. The dataset is significant to shops to determine whether the coupons they offer are benefi-cial or not, depending on the different characteristics and scenarios of the users. A major problem with this dataset was that the dataset was imbalanced and mixed with missing values. Handling the missing values and imbalanced class problems could affect the prediction results. In the paper, we analysed the impact of four different imputation techniques (Frequent value, mean, KNN, MICE) to replace the missing values and use them to create prediction mod-els. As for models, we applied six classifier algorithms (Naive Bayes, Deep Learning, Logistic Regression, Decision Tree, Random Forest, and Gradient Boosted Tree). This paper aims to analyse the impact of the imputation techniques on the dataset alongside the outcomes of the classifiers to find the most accurate model among them. So that shops or stores that offer coupons or vouchers would get a real idea about their target customers. From our research, we found out that KNN imputation with Deep Learning classifier gave the most accurate outcome for prediction and false-negative rate.