Research On Advertising Recommended Systems Based On Filed-aware Factorization Machines

Online advertising is also known as online advertising and Internet advertising.It means advertising on the Internet media.The online advertising has developed rapidly in the past more than 20 years,and has formed a mature and complete industrial process with the population as the target,the core driving force of large data technology and the product oriented delivery mode.The application of big data is the advertisement recommendation system.The main application method is to predict the click rate by the machine learning model.The more accurate the click rate is estimated,the more accurate the advertisement is,the higher the real click rate is,the higher the profit is.The prediction of click rate in online advertising recommendation system is essentially a machine learning pattern recognition problem.The performance and performance of algorithm models are very important in advertising recommendation.Field-aware Factorization Machine is a new nonlinear model of hidden vector learning,which is proposed in 2016.Field-aware Factorization Machine has excellent learning ability for complex features in sparse conditions,and has high application value.Therefore,this paper applies the Field-aware Factorization Machine model to solve the problem of click rate prediction in advertising recommendation system.In this paper,the Avazu data set is used to carry out the steps of filling rate investigation,characteristic frequency statistics,traversing information investigation,data type and coding check to ensure the availability of data.Then the corresponding feature processing is carried out on different features.One-hot encoding is carried out for the category features,and the characteristics of continuous numerical type are carried out.After dividing the bucket into equal intervals,it is discretized,and according to the important discrete value fields,the statistical characteristics of counting and frequency are aggregated.For missing class ID class features,double fields and ID are filled.In this paper,a comparative experiment is carried out from three angles: from the angle of different parameters,for Field-aware Factorization Machine,this paper specializes in constructing three domain characteristics with "user","advertisement" and "login context information" as a domain,and adjusting the more important regular term coefficient L and the number of hidden vectors,K.To verify the performance of the model under different parameters.From the angle of different algorithms,for Logistic Regression,Support Vector Machine,Factorization Machine and Field-aware Factorization Machine model,the discrete value characteristics and continuous value characteristics after processing are input.For integrated learning Random Forest and gradient descent decision tree model,the frequency series will be used in this paper.The lower discrete values and continuous values are used to input.From the perspective of different evaluation indicators,this paper also verifies the performance of multiple models at the accuracy rate,logarithmic loss and AUC value.This paper uses the Python language programming,and calls the tools such as Scikitlearn,XGBoost,Light-GBM and other tools to carry out the experiments of Logistic Regression,Support Vector Machine,Random Forest.The solution is carried out.Experiments show that Field-aware Factorization Machine has achieved better results under multi model and multi evaluation indexes.