Research On Advertising Click-Through Rate Prediction Model Based On Feature Interaction Learning

Advertising revenue is the main source of income for Internet companies.Pushing users their favorite advertisements can not only increase the revenue of advertisers,but also bring users a good advertising experience and increase platform traffic.Clickthrough rate(CTR)prediction is an indispensable part of online advertising push,which aims to improve the accuracy of advertising push.In recent years,the CTR prediction model based on deep learning has achieved good results,but models still have problems such as insufficient feature interaction and high model time complexity.In order to solve these problems,this dissertation proposes two novel CTR prediction models.The main research contents are as follows:(1)Aiming at the problem of insufficient feature interaction,this dissertation proposes a model named Attention-based Feature Interaction Deep Factorization Machine(AFI-Deep FM).The model first leverages the convolutional neural network to learn the potential influence of neighbor features,and dynamically learns importance of the feature itself by the squeeze-excitation network,aiming to enrich feature expression information and improve the quality of feature expression.Then it applies Kronecker product to learn feature interaction,aiming to increase the depth of feature interaction and improve the effect of feature interaction.(2)Aiming at the problem of high model time complexity,this dissertation proposes a model named Domain-based Feature Interaction Learning via Attention Network(DFILAN).The model first groups the features according to the correlation between features and the label,so as to realize independent interaction within and between groups,and improves the efficiency of feature interaction.Then it learns feature interaction from the perspective of feature embedding vector dimension,and uses attention network to learn the importance of feature interaction,aiming to improve the effect of feature interaction.(3)A large number of experiments on the Avazu and Criteo datasets show that compared with the latest model,AFI-Deep FM improves AUC indicator by 0.14% and0.1827%,respectively,indicating that it can better learn feature interaction;DFILAN conducts a large number of experiments on the Avazu and Movie Lens datasets.Compared with the latest model,the AUC indicator is increased by 0.4101% and0.3812%,respectively.At the same time,the DFILAN model parameters are greatly reduced,indicating that it can not only improve the accuracy of CTR prediction,but also can effectively reduce the time complexity of the model,which is more conducive to deployment in practical applications.