Research On Click-Through-Rate Prediction Model Based On Attention Mechanism

As mobile Internet and cloud computing technology develops rapidly,there are huge commercial values behind people’s browsing of various information through mobile Internet,among which the forecast of click rate of advertisement plays a very important role in the realization of advertising business.One of the main factors behind the great success of advertisement delivery is the Click Through Rate(CTR)prediction technology.As the key problem of advertisement recommendation system and online advertising and other commercial applications,CTR prediction plays an important role in the whole mobile Internet.Feature selection is a common dimensionality reduction method,evolutionary algorithms based feature selection method is widely used in solving high dimensional problems.The feature data in click-through rate prediction is usually characterized by high dimensionality and sparsity,which affects the accuracy of click-through probability calculation to a certain extent.Starting from data preprocessing,this paper adopts the feature selection method based on evolutionary algorithm(GACFS)for feature selection of discrete data.In this method,the classification accuracy of the classifier is used as the objective function,the genetic algorithm is used to search the feature combination space,and the elite selection strategy is introduced into the selection operator,so as to ensure that the optimal individuals in each generation can be retained to the next generation.In order to verify the performance of GACFS method,two UCI datasets are selected for comparative experiments,the research results show that the method has better performance than the other two comparison algorithms.On the basis of data dimensionality reduction using GACFS method,aiming at the problem of feature mining in CTR prediction model,this paper proposes a CTR prediction model based on feature selection and improved SENet(FSISC).Firstly,FSISC model selects a better feature subset in feature selection layer.Then the feature subset is embedded in the embedding layer.Then,batch normalization is introduced in ISE module to improve SENet,so that it can quickly learn the new features with the allocated weights.Meanwhile,in IDCN module,the crossover network and neural network are serial combined to achieve arbitrary high-order feature combination.According to the comparison results,the model has good stability and prediction accuracy.