Research On Intelligent Recommendation Algorithm Based On Factorization Machine

In the information age,the complex data and information faced by all walks of life make people's requirements for obtaining effective information more and more urgent.As a set of effective information filtering mechanism,recommendation system can meet people's personalized information needs,so it has been widely used.However,the existing large-scale recommendation system still has two problems to be solved urgently.One is that the inherent characteristic attributes of users and the characteristic information of items in the system are not fully utilized,resulting in low prediction accuracy and cold start problems in the recommendation system.The second is that in the ranking stage of the recommender system,multi-domain category data is often used in the click-through rate(CTR)estimation of advertisements,but the strength of the interaction between the characteristic domains cannot be accurately described,resulting in the model The effective information behind the data cannot be effectively obtained,and the effect of CTR estimation is not good.In the existing CTR prediction models based on deep learning,there is a problem that the system relies too much on deep neural networks(Deep Neural Network,DNN)to learn the hidden information between low-level and high-level features,which causes the burden of DNN to be too large.Affect the overall performance of the system.In response to the above problems,this article has done the following work:1)In the recall phase of the recommender system,this article aims at the cold start problem in the shopping recommendation system.Based on the Probabilistic Matrix Factorization(PMF)model,it makes full use of the user's inherent attribute characteristics and product label information to calculate users separately.Based on the similarity between attributes,more refined user preference similarity,and product similarity,an improved PMF model PSPMF is proposed.Finally,a comparative experiment proves that the performance of the model proposed in this thesis is better than the PMF model.2)In the ranking stage of the recommendation system,this article first addresses the problem of insufficient expression freedom of the interaction between feature domain pairs in CTR estimation.Based on the Factorization Machines(FM)model,a symmetry matrix is introduced.To fully express the strength of interaction between feature domains and propose the FHWFM model.Secondly,the weight of the linear part of the FHWFM model is optimized,so that the model can capture more information contained in the firstorder features,thereby improving the performance of the CTR prediction system.3)In the ranking stage of the recommendation system,this thesis aims at the insufficient capture of feature information in the deep-level CTR prediction model and the excessive learning burden of the DNN part.Try to improve the Deep FM model and replace the FM module of the Deep FM model with the FHWFM module.A Pooling layer is introduced between the Vector Embedding layer and the DNN layer,and combined with the DNN module,the Deep FHWFM model is proposed.It is verified by experiments that the model proposed in this thesis slightly improves the overall performance of the system.