Research On The Application Of Adaptive Feature Extraction Methods In The Click-Through-Rate Prediction

CTR prediction is to analyze whether users will generate clicks on this click object by obtaining data on the characteristics and interactions between users and the object to be clicked.CTR prediction has a wide range of application scenarios and is considered a core issue in popular fields such as e-commerce and recommendation systems,and has also shown great business value.The main challenge of CTR prediction comes from the fact that in practical business scenarios,the features of information content are highly sparse and have a wide range of sources.This makes it difficult to guarantee the efficiency and accuracy of the traditional classification problem in CTR estimation.In recent years,the study of correlation between features and adaptive feature extraction has become a hot issue in the field of deep learning,which also provides new ideas for the study of CTR prediction problem.In this paper,we optimize the feature processing in the CTR prediction process by studying the CTR prediction mechanism,and propose two adaptive feature extraction models by combining clustering algorithms and convolutional neural networks.Model 1 calculates the correlation between samples by clustering algorithm and completes the generation of adaptive features with the help of tree structure;Model 2 calculates the correlation between local features with the help of convolutional neural network,and then captures global features and outputs adaptive features.In addition,the new features generated by both models will be combined with the original features and input to the classification model to achieve a complete set of click-through rate prediction mechanism.To verify the effectiveness of the proposed model in this paper,we use various classification prediction models,including logistic regression(LR),factorization machine(FM),gradient boosted regression tree(GBDT),and multilayer perceptron(MLP),on two public datasets,including MovieLens and Cretio,to compare the prediction performance of the two cases before and after adding the proposed features.The results show that both feature extraction models proposed in this paper are able to extract key information and generate adaptive features that bring improvement in prediction results.Specifically,the main work of this paper is as follows.(1)An adaptive feature extraction model is investigated based on a clustering algorithm and a tree structure.The correlation between features is calculated using a clustering algorithm,and then a tree is generated based on the correlation between samples,and new features are extracted from this adaptively generated tree by formulating coding rules.(2)An adaptive feature extraction model is proposed with the help of convolutional neural network.The generation of adaptive features is accomplished by completing the computation of higher-order crossover between adjacent features and the capture and output of global features by convolutional neural networks.(3)The original features are combined with the adaptive features studied in this paper as input and passed into the classification prediction model to form a complete set of CTR prediction mechanism.(4)To evaluate the two adaptive feature extraction techniques proposed in this paper,in the experimental part we use several classification models,including LR,FM,GBDT and MLP,on two open source datasets,including MovieLens dataset and Cretio dataset,to calculate the performance of the two inputs before and after adding these features,respectively.In addition,the sensitivity of the key parameters of the proposed models is analyzed in this paper.