Research On Click-through Rate Prediction Model Based On Attention Mechanism

As the core research direction of recommendation systems,online advertising and other fields,click-through rate prediction has received extensive attention from academia and industry.In recent years,with the popularity of deep learning technology,many Internet companies and related research institutions have made a series of excellent research results in the direction of combining deep learning with traditional click-through rate prediction models.This thesis summarizes and analyzes the existing click-through rate prediction models,and finds that most models have defects and deficiencies in the cross-feature extraction mechanism,especially the mining of correlation information and importance weights between features is not considered.In response to the above problems,according to the data characteristics of click-through rate prediction,this thesis innovatively proposes multi-state attention mechanism that can efficiently mine the correlation between features based on the self-attention mechanism and hierarchical attention mechanism widely used in the field of natural language processing.Then,combined with the model structure of recommendation system,natural language processing and other fields,this thesis proposes a multi-state attention network which can explicitly construct cross-features based on multi-state attention mechanism.At the same time,based on the design idea of the width and depth joint learning model in the field of recommendation system,and combining the advantages of multi-state attention mechanism and self-attention mechanism,this thesis proposes a multi-state attention and self attention joint learning network model.On this basis,two deep click-through rate prediction models are constructed by introducing deep neural network.The models proposed in this thesis are verified by comparative experiments on a variety of data sets,and have certain advantages in terms of model performance,computational cost,and interpretability.The experimental results show that the clickthrough rate prediction models proposed in this thesis are obviously ahead of other control models in performance;in terms of computing cost,the proposed models have less parameters,faster running speed,and have a great improvement in overall computing efficiency compared with the contrast models;and the models in this thesis have better interpretability and are suitable for the actual production environment.