Research On Machine Learning Algorithm For Recommendation System

With the explosion of information on the Internet,Internet platforms are increasingly relying on recommendation systems to recommend goods to users.The recommendation system can help the platform to filter the appropriate content from the huge amount of information and push it to the user,which on the one hand helps the merchants to capture the interest of the user,and on the other hand can also meet the personalized needs of the user.The recommendation algorithm is the core technology in the recommendation system.The research of the recommendation algorithm is of great significance for promoting the effective application of the recommendation system in the Internet platform.This paper proposes a new click-through rate prediction recommendation model and ranking learning recommendation algorithm for the problems in the two recommendation scenarios of click-through rate prediction recommendation and ranking learning recommendation.The specific research contents are as follows:(1)For the click-through rate prediction recommendation scenario,this study found that the existing model faces three problems: it is difficult to satisfy the memory and generalization functions at the same time;it is difficult to fully mine and combine low-order features;deep model parameters are difficult to learn under the high-dimensional sparse data set of the recommendation system and the recommendation results are too general.In response to these problems,this paper proposes a deep click rate prediction model based on gradient boosting tree and factorization machine.This model combines gradient boosting tree,factorization machine and deep neural network.which can realize both memory and generalization functions,can fully mine low-order feature information and automatically combine low-order features,can learn model parameters in high-dimensional sparse data sets,and the recommendation results are no longer excessively generalized.This paper conducted simulation experiments on the data set of the large-scale competition platform,and compared the related click-through rate prediction recommendation models.Experiments show that this model has achieved better results on the two evaluation indexes of AUC(Area under ROC curve)and Logloss.(2)To solve the problem that the NDCG evaluation index in the ranking learningrecommendation scenario cannot be directly used as the loss function of ranking learning,which leads to the inconsistency of the learning objective of the ranking model and the evaluation index,this paper proposes a ranking learning recommendation algorithm based on approximate function,this algorithm can directly optimize the normalized discount cumulative gains evaluation index and ensure the consistency of the learning objective of the ranking model and the evaluation index.In order to verify the effectiveness of this algorithm,this paper conducted simulation experiments on international public data sets and compared related ranking learning algorithms.Experiments show that the algorithm proposed in this paper has achieved a better evaluation effect on the normalized discount cumulative gains evaluation index.