Research Of Click-Through-Rate Prediction Based On Co-occurrence Relation Network And Deep Neural Network

The Ad Click-Through-Rate(CTR)prediction model is used to estimate the probability that a user will click on an Ad.The CTR prediction model has important commercial value,which directly affects the revenue of the online advertising business.Improving the accuracy of CTR prediction models means increasing company's revenue.The main research works of the thesis are as follows:(1)Current feature extraction algorithms commonly used in the CTR prediction are not fully utilized of collaborative behavior information between users.This paper proposes a feature extraction algorithm based on Co-occurrence Relation Network(CRN).The algorithm can construct an undirected weighted network by using the historical behavior data of the user.The weights in the network reflect the collaborative behavior information between users.Network embedding can then extract the feature representation of the nodes in the network.CRN can be optimized in weight.Therefore,this paper further proposes a feature extraction algorithm based on probabilistic weights.The Co-occurrence Relation Network based on Probability Weights(p-CRN)transforms the undirected weighted network into a directed weighted network,making the information contained in the network more abundant and accurate.The experimental results show that p-CRN brings higher accuracy than the feature extraction algorithms commonly used in CTR predction.(2)Dense features are usually directly connected to a part of the CTR predction model,which does not make full use of the features extracted by the p-CRN.This paper proposes a module,Convolutional Neural Network-based Memory Network(CNN-MN).After combining the p-CRN,the module can learn the feature of the Ad of the user from the user's recent click behavior.The experimental results show that some popular deep CTR models can effectively improve the accuracy of the prediction after using this module.