Research On Advertising Click-through Rate Prediction Model Based On Deep Learning

Advertising click-through rate prediction is a more successful application in the field of recommendation systems.Improving the accuracy of advertising click-through rate can not only make the user experience better but also give more benefits to advertising platforms and advertisers.Although the click-through rate prediction model using deep learning has achieved effective results,there are still some problems,such as the lack of diversity of feature interaction and feature interaction in high-dimensional space.To solve these problems,this dissertation proposes two novel models.The specific research contents are as follows:To enhance the diversity of feature interaction and further improve the accuracy of advertising click-through rate prediction,this dissertation proposes a model named Exploring Different Interaction among Features for CTR Prediction(EDIF).Firstly,the model learns several different embedding vectors for each feature in the embedding layer to enhance the correlation between features;Secondly,the model uses compression to stimulate the importance of dynamic learning feature interaction in the network layer;Finally,the model also introduces the second-order interaction layer and explicit high-order interaction layer in parallel,and learns a variety of feature interactions at the same time.Although the EDIF model fully learns the low-order feature interaction and nonlinear high-order interaction,it ignores the interaction between different dimensions of the embedded vector.Therefore,this dissertation proposes a model named Combining Feature Interaction and Channel Attention for Click-Through Rate Prediction(Fi CANET).Firstly,the model carries out the outer product of the embedded vector to fully learn the interactive relationship between the different dimensions of the embedded vector;Secondly,the channel attention mechanism layer in the model can learn the local information of features and the feature interaction in high-dimensional space through convolution operation;Finally,the high-order interaction layer in the model makes the model learn the deeper interaction relationship of features.In this dissertation,a large number of experiments on the EDIF model and Fi CANET model are carried out on the two public datasets of Avazu and Criteo.The results demonstrate that the model effect of this dissertation has great advantages over the latest model.