Prediction Of Ads’ Click Through Rate Based On Recurrent Neural Network

Along with the rapid development of Internet, in the past few decades, online advertising also get rapid development. Advertising click-through rates(CTR) as an important content of computational advertising, start to become an essential part of the industry. Advertiser use machine learning algorithms, based on a lot of historical data for model training estimate ads’ click through rate and accurate ads’ click through rate estimate can make advertising more targeted and efficient, which will also improve the real hits and increase income. Even though the linear model can be simply estimated ads’ click through rate, but the learning ability of linear model is limited, not only it will reduce the efficiency when more and more data characteristics disappear, but also excessive fitting will be easily happened in the learning process,which will results in reducing the effect of training. This method is based on neural network algorithm with nonlinear excitation function and multilayer node structure.Because this method can provide deeper analysis for the complex relationship between a large number of nonlinear characteristics, so as to improve the estimate ability of the model. Among them, the recurrent neural network is a network which consist of ring structure, can store the output of the previous moment for neurons.Also, it are equipped with the neural network which has strong ability of optimization calculation. In this thesis, the main work includes the following three aspects:(1) Based on different models for the characteristics of the corresponding processing, logistic regression model utilize the matching dominant feature to extract the hidden user attributes, and then the original characteristics of the different types of conversion are hashed into the same type of eigenvalue. Random forests model uses the established feature dictionary, filtering the sample frequency which is extremely low, then processing characteristics by using one- hot coding.Based on the Model ofneural network, this thesis first calculates the frequency characteristics, as well as establishes the characteristic frequency of the dictionary, then converts the character features into an integer. In the end, the dispersion characteristics are standardized, and each features of value are changed into the range of [0, 1].(2) Although the recurrent neural network model has been applied to the predict of ads’ click through rate,this model results in problems using gradient descent. When approaching the minimum gradient, this model may arise the gradient disappear problems, which affects the model predict results. In this thesis, the recurrent neural network based on LSTM(long short term memory) are used to forecast ads’ click through rate. In order to prevent the problems of the gradient disappear, LSTM structure are used to modify RNN.The experimental results show that the recurrent neural network model based on LSTM has obtained the good effect on forecasting ads’ click through rate.(3) In this thesis, logic regression model is utilized with python language. And a random prediction model, BP(Back Propagation)neural network model, the recurrent neural network model and the recurrent neural network model based on LSTM are built based on the python language toolkit.In this thesis the sigmoid function and ReLu function are adopted to train the recurrent neural network. As the experiment proved, ReLu function converges faster, and has better result of model prediction.Logloss method is used for model evaluation.Compared with the AUC(Area Under roc Curve), logloss can reflect the more accuracy of model estimating ads’ click through rate.