Improvement Of DeepFM Algorithm Based On Matching And Correlating Subnets

In the digital age,the identity and behavior of each user can be quantified by numbers,and there are countless commercial values behind the numbers.According to the quantitative figures,each platform will accurately deliver the content that can attract users to watch or the advertisements that users are interested in.A good click-through rate will not only bring considerable benefits to the platform side,but also bring users a comfortable experience.Therefore,the research on the prediction of click-through rate is of great significance.Firstly,this thesis describes and analyzes the structure,advantages and disadvantages of two mainstream click-through rate prediction models,and on this basis,proposes a new fusion model.This thesis analyzes and explains the mainstream click-through rate prediction models,including a neural network model based on factorization machine and a neural network model containing three sub-networks.Inspired by the above two models,this thesis proposes a new fusion model,DeepFMF model（Factorization-Machine based neural network with matching subnet）.The new model can learn the relationship between features,specifically the similarity between user features and advertising features,and also learn the high-order combination of features to improve the memory ability of the model,so it can improve the performance of the model.Secondly,on the basis of the newly proposed click rate prediction model,DeepFMF,this thesis continues to improve and proposes a more advanced fusion model,DeepFMC（Factorization-Machine based neural network with correlation subnet）.The model not only has a matching sub-network that can learn the similarity between user features and advertisement features,but also has an association sub-network that can learn the degree of association between advertisements.The sub-network introduces the sliding window in natural language processing,sets the window for the advertisements in the user’s advertisement click history,and adopts the negative sampling technology in the window,so that the network can better learn the correlation degree between advertisements.Finally,the thesis simulates and compares the new model,analyzes the influence of different factors on the model,and compares it with other models.Two commonly used indexes in the field of experimental click-through rate prediction are used to evaluate the performance of the model,and the influences of different factors such as super-parameters,dropout ratio and the number of hidden layers on the performance of the model are analyzed.Moreover,a comparative experiment is carried out on two data sets,and the new model proposed in this thesis is compared with the mainstream click-through rate prediction model.The experimental results show that the sub-network learning the relationship between features can greatly improve the prediction performance of the model in all indexes.Therefore,it is proved from practice that learning the relationship between features will have an impact on the improvement of model performance,and it is also proved that the two new models are feasible and superior.