Research On Prediction Method Of Ads' Click-through-rate Based On Neural Network

With the rapid development of modern information technology,online advertising in industry has risen rapidly.As a new form of advertising,it has huge market potential and commercial value,and it has been widely watched.Click-through-rate(CTR)estimating is one of the key issues to Internet advertising.By using machine learning and big data technology,to a large extent,the data in the advertisement about system log is predicted.The current main method is to estimate the click-through rate of advertisements based on historical data.With efficient and accurate click-through estimation of advertisements,advertisements are been placed more accurately,and the actual click-through rate about users is increased,so that advertisers' revenues are greatly increased.In this thesis,we mainly use the click-through rate based on the neural network as the research object.At the same time,we use the click-rate estimation algorithm based logistic regression as the baseline algorithm to analyze the difficulties and problems that currently exists in both advertising and baseline algorithms in different situations.Put ing forward the corresponding optimization method.(1)In terms of display advertising,cost should be taken into consideration.At the same time,the cost of feature processing should be minimized and the accuracy of model prediction should be comprehensively improved.Then a universal model should be designed,ie,an innovative model based on FM algorithm and neural network.Instead of adopting the traditional one-hot encoding method,the vector product is used to embed the high-numbered feature vectors into the low-dimensional vector space.Based on this,the neural network model is applied to acquire better performance in Classification of field data.The experimental results show that the click-rate estimation model based on factorization machine and neural network used in this thesis can obtain better prediction results than mainstream methods.(2)In terms of search advertising,based on the understanding of customers' need and searching habits,this thesis designs a model based on deep neural networks and adopts a dropout method to reduce feature fitting.During the training process,randomly selects some hidden layer nodes according to a certain probability.Their weights do not participate in training during training;GPU-based block-based schemes are used to increase the speed of operations.This thesis optimized the traditional online advertising CTR estimation algorithm,further improving the click-through rate,and making the click-through rate about prediction result more credible.The experimental results show that the dropout method and the GPU-based blocking scheme reduce the fitting phenomenon and improve the operation speed to some extent.