Research On Click Rate Prediction Algorithm Based On Feature Interactio

With the rapid development of deep learning and big data technology,an increasing number of Internet companies are focusing on how to improve the prediction accuracy of ad click-through rates.The higher the accuracy of click-through rate prediction is,the more accurate the information recommended to users becomes,and the higher the additional revenue for advertisers would be.To improve the accuracy of click-through rate prediction,feature interaction is often performed by manual combination.However,this method is costly,inefficient,and does not achieve high accuracy,so it is an important task for click-through rate prediction algorithms to accurately and efficiently explore the implicit cross features in the sample information to improve the prediction accuracy.Address this problem,this paper conducts a series of analyses and research on the mechanism of modeling cross features in frontier prediction algorithms.The specific work has two aspects as follows:(1)The mechanism of feature interaction in frontier algorithms for click-through rate prediction is investigated.Firstly,to address the problem of unclear crossover rules and unknown information limits in the current public datasets,we build experimental datasets with only low-order and high-order crossover feature rules.Next,we test cutting-edge click-through rate prediction algorithms on these datasets,including GBDT and deep neural networks,and analyze and explain the experimental results in terms of performance,principle and interpretability.The experiments finally show that most algorithms are difficult to find the optimal solution for feature interaction,especially decision tree models such as GBDT,which are good at modeling correlations but are less noise-resistant and not suitable for mining interaction features.In addition,the accuracy of click-through rate prediction can be significantly improved by deep learning algorithms.(2)To address the problem of insufficient feature interaction expression in current prediction algorithms,we propose a fusion feature interaction and selection model named XgbAFM.The core idea of this model is to use XGBoost to automate feature engineering and filter a certain proportion of features into the deep neural networks according to feature importance;meanwhile,to perform deeper and more adequate feature interaction on the original data,we put the character-based features after Embedding into the FM and DNN respectively,and use the attention network to dynamically learn the importance of the crossover features,which take into account the low-order and high-order feature representations.Experiments on the Criteo dataset show that the XgbAFM model effectively improves the prediction accuracy of ad click-through rates with an AUC increase of 1.6013%compared to the optimal benchmark model.