Feature Engineering of Click-through-rate Prediction for Advertising

We present the problem of click-through-rate (CTR) for search advertising in ALiMaMa, which displays user information, item information, shop information and trade results. Traditionally, people use logistic regression (LR) to predict it. However, because of the lack of learning ability and the sparse feature matrix, the prediction results are always not so satisfying. In this paper, we mainly propose some feature engineering methods based on gradient boosting decision tree (GBDT) and Bayesian smoothing to obtain a wonderful feature, which has more useful information and is not so sparse. Also, we use xgboost (XGB) instead of LR as our prediction model. The proposed methods are evaluated using offline experiments and the experiment results prove that the log loss drop near 5% after using these feature engineering methods and XGB. Obviously, it is an excellent performance.