Click-Through-Rate Prediction Based On Deep Neural Network Model

Online advertising is becoming increasingly important with the fast development of the Internet. One of the most important problems in online advertising is Click-Through-Rate (CTR) prediction which uses large scale data from logs and machine learning methods to build a model, and then use this model to predict the probability of the user clicking the advertisement and then present the user with the advertisements he(she) may probably click. Among these models, Logistic Regression (LR) is a widely used one. However, LR has some intrinsic limitations,1) LR is a linear model which is hard to capture the underlying nonlinearity in the data;2) LR has too many parameters which make it easier to over-fit. These limitations make LR inappropriate in CTR prediction.In this work, I surveyed the common used machine learning methods in CTR prediction and proposed a framework that use Deep Neural Net-work (DNN) to predict CTR. UP to my knowledge, no one had use same method to predict CTR. The main works of this paper are as follows:(1) I do much data analysis and preprocessing work, discretize the numerical features and use one-hot encoding to category some of these features.(2) I implement LR model using Python, and do some experiments on real datasets. The results are used as the baseline to compare with the performance of DNN model. (3) I survey the DNN model and propose the framework to predict CTR based on DNN model and confirm the performance on real datasets using the open source tool Petuum, we select two of sigmoid and relu ac-tivation function, features are specially designed for DNN model.