Research And Application Of Advertising Click-through Rate Prediction Method Based On Feature Interaction

Click-through rate prediction(CTR)is a very important research direction in recommendation system.Click-through rate prediction(CTR)refers to a certain advertisement that needs to be clicked next by the user under certain circumstances.Before clicking,the system predicts the probability that it will be clicked according to the recommendation algorithm.If the AD has a high probability of being clicked,show the user the AD;if the probability is low,don’t show it.Advertising recommendation and placement are essential for many Internet companies.In the Click-Through Rate Prediction(CTR)prediction task in the era of big data,the input data is not only large in quantity but also high in feature dimension,which greatly affects the prediction performance of important feature selection and feature interaction.This paper optimizes and improves the problems existing in feature selection and feature interaction in the click-through rate prediction task.The main research contents include:(1)A click-through rate prediction model based on feature weighting and automatic interaction is proposed.Firstly,aiming at the problem of information loss caused by dimensionality reduction in the selection process of feature importance in the previous click-through rate prediction model,a feature weighting method without dimensionality reduction is proposed,which can effectively avoid information loss by introducing ECANet module to learn the original feature weight.Secondly,in view of the limited modeling ability and poor interpretability of a single model,the multi-head self-attention network and deep neural network(DNN)are respectively used to automatically learn explicit and implicit feature interactions.Finally,the outputs of the two are spliced together and input into the prediction layer for prediction.Experiments on four classical click-through rate data sets,such as Criteo,prove that our proposed weighting method can effectively solve the problem of information loss,and the combination of explicit and implicit feature interaction can enhance the interpretability of the model.(2)A click-through rate prediction model based on the interaction between global attention network and feature is proposed.For most models,the learned low-order and high-order feature information is directly spliced for prediction,while the relationship between the low-order and high-order feature information is ignored.We designed a global attention interaction module(GAN)to further interact the high and low order interaction information learned from the interaction layer and fully explore the important information.In addition,the model uses attention mechanisms to learn the importance of different interactive features and original embedded features.Finally,the off-line experimental results on Criteo,KDD12 and Movie Lens-1M data sets demonstrate that the method can effectively improve the accuracy of click-through rate prediction.(3)A click-through rate prediction system based on feature interaction is designed.The system uses the collected user characteristics to calculate the probability of an advertisement being clicked,so as to help the platform recommend the most suitable advertisement for each user.Accurate recommendation can increase user stickiness and improve the quality of advertisement delivery on the platform.