Feature Selection in Click-Through Rate Prediction Based on Gradient Boosting

Abstract

Click-Through Rate (CTR) prediction is one of the key techniques in computational advertising. At present, CTR prediction is commonly conducted by linear models combined with $L_1$ regularization, which is based on previous feature engineering including feature normalization and cross combination. In this case, the model cannot realize automatic feature learning. This paper uses the ensemble method for reference and proposes a feature selection algorithm based on gradient boosting. The algorithm employs the methods of Gradient Boosting Decision Tree (GBDT) and Logistic Regression (LR), and further conducts a positive analysis in the data set of kaggle-CTR prediction on display ads. The experimental result verifies the feasibility and validity of feature selection method. Moreover, it improves the performance of CTR prediction model, whose AUC value reaches 0.908.