Click-Through-Rate Estimation Based On Deep Learning

Click-through rates(CTR)estimation is the core module of the advertising system,which has a large number of applications in search engines,video sites,shopping sites,life services and other large sites.Accurate estimation plays a important role to enhance the marketing effectiveness of advertisers,increase revenue of advertising platform,improve the user experience.The common application models include logistic regression model and factorization model.Deep learning has strong ability of data fitting,and has made great progress in many fields.In this paper,the deep learning method based on feature field is introduced to deal with the CTR estimation.This paper introduces the application of traditional model in CTR prediction,analyzes some limitations and shortcomings of traditional model,and introduces deep learning method.The data processing in the rate estimation problem usually has high dimensionality and sparse features,so it is difficult to directly input to the neural network,and it is difficult to train effectively in deep learning.By dividing all the features into different feature field,the sparse data inputs of different feature field correspond to different fully connected units.Each unit will be fully connected network mapping input data to a fixed number of dense output neurons,these neurons together with the real input output characteristics of the newly constructed concentration from the data as a whole deep learning network.The model of deep learning based on feature field for experiment in the open Kaggle platform based on data sets,and compared with the logistic regression model and factor decomposition model as the benchmark model,the experimental results show that the characteristics of domain of deep learning is better than the benchmark model based on can be effectively applied to the CTR problem..