Website Rating Prediction Based On BiasSVD And Ensemble Learning Model

All kinds of websites have been constantly emerging in recent years,people have entered the era of information overload.In order to solve the problem of information overload,the recommender system emerges at the historic moment and has been widely used in mainstream websites.Rating prediction and Top N recommendation are two major recommendation tasks.This thesis mainly focuses on the study of rating prediction.There are three problems in rating prediction: data sparsity,cold start and scalability.The existence of these problems leads to the high error of traditional models.Aiming at traditional problems of the above models,this thesis proposes a solution combining BiasSVD matrix factorization and the ensemble learning model.The solution mainly has the following three innovations:(1)Combining the most important contextual information: time.On the basis of BiasSVD matrix factorization,user bias term and item bias term are respectively combined with the inverted Sigmoid curve and the relationship curve between item popularity and rating.BiasSVD matrix factorization is transformed into BiasSVD matrix factorization incorporating time.(2)The underlying connection of users and items is mined to create features.Firstly,row vectors of the product of the user latent matrix and the item latent matrix learned by BiasSVD matrix factorization incorporating time are used to obtain the clustering feature with the aid of K-means clustering algorithm.Then,user interest and item popularity can be respectively obtained from user bias term and item bias term as two features.Lastly,rating is added as weight to the graph-based random walk algorithm to obtain the access probability of each item.(3)Features mentioned in the second innovation are taken as deep features and combined with basic features such as the gender of users,the occupation of users and so on.Then these features will be input into Light GBM ensemble learning model to make prediction.The solution proposed in this thesis not only keeps the advantage of matrix factorization that the model is not affected by data sparsity,but also fully mines the underlying connection of users and items to create deep features.In addition,using Light GBM model to combine deep features and basic features alleviates the cold start problem in a certain sense.In the experiment of this thesis,Mean Absolute Error(MAE)and Root Mean Square Error(RMSE)are used as evaluation indexes.The proposed model is compared with UserCF,ItemCF,FunkSVD,BiasSVD and other models.The result of the experiment shows: MAE decreases by 17%-29% and RMSE decreases by 16%-27%.