LAUSR

Besides the ID class features, the advertisement click log file contains many significant features, which make the study of the advertisement clickthrough rate prediction more difficult. In this study, we convert original features into numerical meaningful ones, which reduce the sparsity and redundancy. In order to solve the problem of class imbalance, we propose a downsampling algorithm based on K-means model to classify large samples, then divide them into some sensible and rational features by the heuristic methods. To further improve the feature representation, we finally select and combine features by the Gradient Boosting Decision Tree model and process high-dimensional features by the logistic regression method. We conducted experiments on the dataset of Tencent SOSO and demonstrated that our approach outperforms the state-of-the-art baseline methods by 0.05% on average in terms of R2 and by 50.5% on average in terms of RMSE.