Research On The Click-through Rate Prediction Algorithm Of Online Advertisement Based On Feature Interaction

Online advertising,also known as internet advertising,refers to the advertisements placed on search engines,e-commerce platforms and other online media.With the rapid development of internet technology,online advertising has formed a technology-driven delivery mode aiming at precision delivery in just over twenty years.The core of online advertising profit comes from click-through rate which is the ratio of advertising click and advertising display,each single click can bring revenue to advertisers.In addition,the online advertising system generally ranks the candidate ads according to the click-through rate,and recommends the most interesting ads to users,which can effectively improve the user experience and enhance the viscosity of users.Therefore,the task of click-through rate prediction is very important in the online advertising technology system.At present,the click-through rate prediction task mainly uses semi-structured data such as given user information,advertisement information,and context information to make prediction,and recommends the corresponding revenue of user-related advertising settlement based on the obtained click-through rate.Based on this scenario,user characteristics,advertisement characteristics,and contextual characteristics are important characteristic information in the click-through rate prediction task.In order to improve the accuracy of click-through rate prediction,experts are often used to select some features to interact with each other manually,so as to get richer heterogeneous information,so as to get more accurate click-through rate.However,this method is too inefficient and expensive,and some implicit feature correlations are difficult to find.Therefore,how to effectively conduct feature interaction and fully explore the hidden relevant information between features has become an important research direction of click-through rate prediction method.In recent years,in order to improve the performance of the click-through rate estimation model,the click-through rate prediction algorithm has combined the ability of deep learning to mine implicit features with some ability to display the interaction of learned features,and has made great progress.The ability to learn feature interaction in the click-through rate estimation task becomes the key to improving the prediction performance of the model.In this dissertation,the actual application scenarios and the current research frontier are combined to study the key issues of feature interaction in the current clickthrough rate prediction method,so as to improve the performance of the model and improve the efficiency of the model as much as possible.The specific work mainly includes the following three aspects:(1)Starting from online advertising,this dissertation divides the click-through rate prediction model into machine learning and deep learning,briefly introduces the overall framework and evolution process of the click-through rate prediction model,and explains the internal logic of the model evolution.Finally,the evaluation matrix of the CTR prediction model are introduced and analyzed.(2)Conduct a series of analysis and research on the problem of insufficient feature interaction ability in the current click-through rate prediction method.Firstly,a recurrent interaction network is constructed to solve the explicit feature interaction problem.The recurrent interaction network uses matrix operation to explicitly conduct feature interaction and can learn the nonlinear relationship between feature interactions by using 1×1 convolutional neural network.Recurrent interaction network is also cyclic,which learning interactive features of different orders according to the number of network layers.In order to make the model have the ability of both explicit and implicit feature interaction,this dissertation combines the feature interaction network with the multi-layer perceptron in deep learning,and proposes a deep feature interaction network.Deep recurrent interaction network can simultaneously mine explicit and implicit interaction features,which further improves the overall performance of the model.Finally,experiments are carried out on three real datasets,and the results show the superior performance of the deep recurrent interaction network.(3)Aiming at the noise problem that may be caused by redundant interactive features in current click rate prediction algorithms,this dissertation proposes a click rate prediction model that combines feature interaction and selection.This model proposes an interaction feature selection function based on the three-way decision theory.The interaction feature selection function can strengthen important interactive features to a certain extent,retain important interaction features,and eliminate redundant interaction features.The experimental results on three public data sets show that the click-through rate prediction model that integrates feature interaction and selection is better than the existing model in terms of efficiency and performance.