Application Research Of Factorization Machine In Recommended Fields

The rise of the Internet has brought about a dramatic increase in the amount of information,satisfying users' needs and desires for knowledge.However,with the rapid development of the Internet,the explosion of information has made it impossible for users to directly obtain the content they really need in front of massive data.On the contrary,it has become more difficult to obtain effective information and form information overload.The recommendation system is one of the main means to solve the above problems.Currently,the mainstream model in the recommended field is a collaborative filtering algorithm,which is divided into collaborative filtering based on User and Item and collaborative filtering based on model.In the model-based collaborative filtering algorithm,the matrix decomposition series algorithm is the main branch,but it has drawbacks in mobility.Factorization machines model can simulate common matrix decomposition models by changing the input real value vector form,such as SVD++,PITF,etc.,effectively avoiding the drawbacks of traditional matrix decomposition model which need to define the model expressions and optimization methods separately for each specific task.Therefore,this paper will focus on the series of factorization models.First of all,the user behavior information implied in the user-item scoring matrix reflects the user's interest preference to a certain extent,which is very valuable information for recommendation,but the common collaborative filtering algorithm does not make this information very good use.Aiming at the above problems,this paper proposes a fusion model of word2 vec and factorization machine,which is to model the sequence information through word2 ve technology,adopting its output as the part of input of factorization machine model to further improve the accuracy of the score prediction of the factorizermodel.The above fusion model has been proved by experiments to achieve better returns in reducing the prediction error of the score.Secondly,based on the factorization machine,the field factorization machine proposes the concept of field,which makes it better in feature modeling.But it's usually used in classification domain,and here we use it to solve regression problems like score prediction.In order to further enhance the effect of the model,we also integrate the domain factorization machine model and the Mini-Batch Kmeans++ algorithm to further improve the accuracy of the score prediction of the domain factorizer model.The experimental results demonstrate the effectiveness of the scheme in reducing the prediction error of the score.