Research On CTR Method Based On Neural Network

Our society has entered the information age.People's browsing behavior of web pages and other information is called traffic,which contains great commercial value.Most Internet companies are exploring the means or methods of flow realization.As an important way of commercial realization of Internet companies,advertising click through rate estimation plays an irreplaceable role in commercial realization.Among them,the technology behind the great success of advertising industry is click through rate(CTR),so the research on CTR has great commercial value.In recent years,with the successful implementation of CTR technology,a large number of scholars have also invested in the research of this field.There are many proposed models in this field,such as logical regression,tree based model,factorization machine based model and CTR model based on deep learning.However,many of the current work is to use Hadamard product and inner product and other simple methods to calculate the interaction between features,and does not pay attention to the importance of the features themselves.This paper starts with the method of feature construction.Firstly,the compressed excitation neural network(senet)is introduced to extract the importance weight of a single feature in the whole,and then the attention mechanism is used to give consideration to the interactive feature weight.In this paper,the shallow prediction model(senet-atten CTR)is proposed and applied to the real open source dataset.Experiments show that the model has better prediction ability than the traditional CTR prediction model,and the features constructed by the model have better interpretability,and there is no need to manually mine the relationship between features in the process of feature construction,which greatly improves the efficiency of the task.On this basis,considering the generalization ability of the model and the deep level data mining of the characteristics,the deep neural network is applied to the above CTR prediction model,and the deep-senet-atten CTR is proposed to improve the model,which is compared with the traditional deep CTR prediction model,Experiments show that deep-senet-atten model has better prediction ability than traditional CTR prediction model in real open source data set.