Font Size: a A A

Research On Click-through Rate Prediction For Search Advertising Based On Deep Learning

Posted on:2016-05-11Degree:MasterType:Thesis
Country:ChinaCandidate:S Q LiFull Text:PDF
GTID:2308330479990068Subject:Computer Science and Technology
Abstract/Summary:PDF Full Text Request
With the flourishing development of Internet Technology, it becomes more and more popular for network marketing in the background of big data. As a new form of advertising, online advertising shows a huge market potential and commercial value. Search advertising, which delivers ads according to the user’s search content, is the largest and fastest-growing form of online advertising and has been the main income for Internet Industry. Predicting CTR is the most critical technology for search advertising, for it is not only related to ad rank, but also affects ad click charges. It is a meaningful work to predict CTR correctly based on massive historical search data. At present, most of the existing work predict search CTR based on shallow models. In shallow models, the meanings of features are fixed and isolated, without considering the inner relation between the features.This paper is focused on predicting CTR of search advertising given the specific information. In order to predict the result effectively, a deep neural network is employed to dig information between the different features. The contents could be generalized into three parts:First, after giving the definition of the CTR of search advertising, data distribution is analysised. According to the knowledge of search advertising and characteristic used in practise, six categories of features are extracted based on the processing data. To address the problem of memory limit and time consuming, this paper employs a data-chunked method, which is based on GPU.Secondly, considering the weakness of Na?ve Bayes, Logistic Regression model and Support Vector Regression Model for predicting CTR, we put forward a deep neural network-based method for predicting the CTR of search advertising, what is more, dropout method is used to prevent train from overfitt ing. Based on the same feature set, DNN-based method gives a better performance when compared to main method for predicting CTR of search advertising.Thirdly, a convolution neural network-based method is employed to predict the CTR of search advertising. The operation of convolution and pooling makes it possible for the model to learn the relation between local features and global features. Experiments are conducted on the data set of KDD Cup 2012. The performance shows that CNN-based method make a big improvement for predicting the CTR of search advertising.
Keywords/Search Tags:click-through rate, search advertising, deep learning, deep neural network, convolution neural network
PDF Full Text Request
Related items