Click-Through Rate Prediction Using Feature Engineered Boosting Algorithms

Click-Through Rate (CTR) prediction plays a critical role in online advertisement campaigns and recommendation systems. Most of the state-of-the-art models are based on Factorization Machines and some of these models try to feed mapped field features to a deep learning component for learning users' interests by modelling feature interactions. Deploying a model for CTR is an online task and should be able to perform well with a limited amount of data and time. While these models are very good at prediction inferences and learning feature interactions, their deep component needs a vast amount of data and time and does not perform well in limited situations.In a recent article, a combination of boosting algorithms with deep factorization machines (XDBoost algorithm) has been proposed. In this paper, we use a boosting algorithm for prediction inference with limited raw data and time. We show that with an appropriate feature engineering and fine parameter tuning for a raw boosting model, we can outperform XDBoost method and get better results. We will use exploratory data analysis to extract the main characteristics of the dataset and eliminate the redundant data. Then, by applying grid search scheme, we select the best values for the hyperparameters of our model.