An Efficient CTR Prediction Method Based On Improved DeepFM Algorithm

In the Internet age,advertising is a very important means or method for Internet companies to make profits.As the core research issue of advertising and other businesses,click through rate prediction plays an important role in the Internet field.In the past few years,it has been transformed and driven by big data technology,repeatedly guided and evolved by products,and industrial processes have developed rapidly until they are mature.Online advertising is now a multi billion dollar industry,which has become one of the successful cases of making money in the field of machine learning.In recent years,artificial intelligence and deep learning have made a lot of research progress in various fields of computer,making major Internet enterprises and scientific research institutions begin to explore the role of deep learning technology in the scenario of advertising click rate prediction,and have made a lot of commendable results.With the advancement of cloud computing technology,various products in the whole Internet field have led to a blowout like growth of data scale.With the development of various data mining and machine learning technologies,these technologies can accurately and efficiently construct models from large-scale data sets,and dig out valuable information for companies or society.Sponsored search advertising,advertising context,display advertising and real-time auction all rely heavily on the learning ability of the models,which can accurately,quickly and reliably predict the advertising click through rate.But this problem also leads to the problem of scale,which was almost unimaginable even 10 years ago.Typical industrial models can use the corresponding large feature space to predict billions of events every day,and then learn from a large number of data.For the huge training data,the scalability of the model is greatly limited,many useful features or more complex models can not be used,so a variety of sampling strategies are derived.The influence of different scale data on model selection is huge.The traditional matrix decomposition method has been widely used in prediction,recommendation system and other fields.The main reason is that the performance of the algorithm can meet some business needs and the algorithm has high scalability.However,in the face of big data environment,people can get more context information,and the traditional matrix decomposition method is lack ofcontext information utilization.In the face of this challenge,a factor decomposition machine model is proposed and widely used.As the prediction of ad click rate is an indicator to measure the possibility of ad being clicked by users,the ultimate purpose of the prediction of ad click rate is to improve the search experience of users and the exposure rate of ads to potential customers to improve revenue,so as to achieve win-win results.The traditional advertising click rate prediction model has a low accuracy,but also needs to construct a large number of artificial features,which consumes a lot of time and human resources.In order to solve this problem,based on the factorizer and its optimized model,this paper designs a model which has good prediction accuracy and does not need to be constructed manually.In the case of high sparsity,the factorizer can learn the low-order features,while the deep neural network can learn the high-order features,and fuse the lifting tree to enhance the accuracy and robustness.Experiments show that the model designed in this paper has higher accuracy and availability than the factorizer and its derivative models or traditional models,and has the practical ability to solve the problem of advertising click through rate prediction.No matter how to improve the factorizer model structure or how to improve the accuracy of CTR prediction model,the research and exploration in this paper are practical meaningful.