Research On Click-through Rate Prediction For Online Advertising Based On Storm And Online Learning

As a new advertising and marketing model,online advertising greatly prompts the development of the advertising industry.In the model of online advertising marketing,the CTR(click through rate)will directly affect the interests of the advertisers and the web publishers.Thus,the prediction of the CTR is important to the online advertising industry.The core of predicting the CTR is machine learning,which has made a technological breakthrough in recent years.For the problem of machine learning,many researches have proposed a number of algorithms,such as logistic regression algorithm,factorization machines and so on.And these models can be solved by some update algorithms such as stochastic gradient descent,FTRL-Proximal and so on.Based on the advantages and disadvantages of these algorithms,a reasonable improvement can solve the problem of machine learning with a better result.Besides,there exists another method of processing such large-scale data of advertisement information,online machine learning.The principle of it is to complete machine learning in a distributed system.The main work of this thesis is as follows:(1)Based on the stream processing,this thesis has summed up some efficient method of reducing dimension and filtering features in processing ad logs.(2)This thesis has introduced the logistic regression algorithm and factorization machines.Then combined with the FTRL-Proximal algorithm,this thesis promotes an improved algorithm based on factorization machines.And experimental results proof that this improved algorithm has better performance and sparsity.(3)Thesis makes a presentation of several popular distributed stochastic gradient descend methods.Making use of the principle of parameter server,a real-time online machine learning system based on Storm has been designed in order to meet the demand of online processing and training for the continuously increasing ad data.