Research On The Algorithm Of Click-Through Rate Prediction

Flyers,newspaper publications,and television placements are a thing of the past.With the advent of the 5G era,Internet advertising has long dominated.How to make advertisers successfully promote,the media platform shows intelligence,users are willing to accept,and push the right ads to the right users.Get accurate CTR estimates,sort them,deliver based on results,and estimate again later.In such a business closed loop,it is indispensable to know the predicted value of click-through rate in advance.Advertising monetization is one of the means by which the major Internet giants are lazy to survive.No matter in academia or industry,research on related algorithms of click-through rate prediction is very popular.Because of the commercial value and availability of this issue,researchers are happy with it.In this paper,from the standpoint of a media platform,the model experiment effect improves the prediction performance of the click rate while ensuring the smallest possible loss.The specific main research includes the following four aspects:First,advertising-related data is inherently sparse.Improper feature selection not only has the risk of overfitting but also dimensional explosion and the possibility of memory overflow.In this thesis,after careful data analysis,the advertising effect generation process is used as the basis for constructing business characteristics.By training features with smaller dimensions,better prediction results can be obtained.Second,in actual business scenarios,the emergence of new users may occur at any time.This will cause the user to cold start.Solutions to such problems are rarely involved in calculating advertising.This thesis improves demographic informatics based methods in recommendation systems.From the two dimensions of positive and negative,construct the characteristics table of each clicked or unclicked crowd of advertisements,compare the characteristic attributes of new users with them,and calculate the rough click probability value,which is used as the cold start related feature.Experimental results show that this feature can effectively improve the prediction effect of each comparative experimental model.Third,read the literature on the algorithms of the classic model and draw on the ideas of previous generations.This paper proposes a new network structure PdeepNFM model.Inspired by the Wide&Deep,the model is a "depth + depth" parallel network structure.Dimension reduction is implemented from the input layer to the Embedding layer.After inputting to the Bi-Interaction layer and the Product layer,and entering the respective hidden layers,relatively low-order and high-order features are constructed.Map by Sigmoid function,and output click-through rate prediction value.Both sides of the model are trained jointly,and the features complement each other.The experimental results show that the new model proposed in this paper is superior to other comparative experimental models in the two evaluation indicators of AUC and Logloss.Fourth,verify that the prediction performance of the new model under different optimizers,activation functions,and Dropout retention ratios are in line with basic theoretical characteristics.On this basis,the continuous features used in the model are sent to the gradient boosting decision tree for training,and the split features and the original category features are stitched into the new model.Experimental results show that the optimization method can further increase AUC by1.84% and reduce Logloss by 0.85%.