Deep Click Prediction

Online advertising is one of the most effective ways for businesses of all sizes to expand their reach, find new customers, and diversify their revenue streams. Predicting whether a user clicks on an impression is a central problem in display advertising. Display advertising involves features regarding contexts (publishers) as well as users and ads (advertisers). This implies more complicated interactions between these features, which should be taken into consideration when clicks are predicted. However, traditional approaches, such as linear models are weak in modeling these complicated interactions between context, users and ads. We propose three deep models, Multilayer Perceptron, Convolutional Neural Network and Autoencoded Gradient Boosted Decision Trees, and compare the results to the traditional models, Logistic Regression and Gradient Boosted Decision Trees, prevalent in the industries. We also treat this classification problem as an Anomaly Detection problem by varying the skewness ratio of the two classes. We were able to marginally beat the traditional models because of the reasons we discuss later.