Font Size: a A A

Design And Implementation Of A New Hybrid Recommendation Model

Posted on:2020-07-09Degree:MasterType:Thesis
Country:ChinaCandidate:Z K GuoFull Text:PDF
GTID:2428330572973556Subject:Computer technology
Abstract/Summary:PDF Full Text Request
With the rapid development of Internet technology,we have entered the era of information overload.Faced with the huge amount of information on the In ternet,it takes a lot of time for people to find the part of information they really need.In this context,recommendation system emerged as the times require.The recommendation system intelligently recommends information that users may be interested in by analyzing their past behavior.After nearly 20 years of development,recommendation system has been used in many enterprises in the actual system,such as Yotube,Amazon,Taobao and today's headlines.Although recommendation system has been widely used in enterprises,it still faces the problems of accuracy,cold start and scalability.Collaborative filtering algorithm is the most common and successful algorithm in recommendation system.Matrix decomposition is a widely used collaborative filtering model.Matrix decomposition model updates the parameters by stochastic gradient descent method(SGD).It needs to traverse the training set several times in the process of model training.The matrix decomposition model based on SDG is an off-line algorithm,which needs to construct an off-line training set,and its scalability is limited,that is,with the increase of data sets,training time increases sharply.In order to solve the scalability problem of matrix decomposition model,this paper proposes a matrix decomposition model based on FTRL.It updates the parameters by FTRL method,and only needs to traverse the training set once,which greatly reduces the training time of the model.At the same time,it is also an online learning algorithm,which can update the model in real time according to the user's feedback data,so as to respond quickly to the user's behavior in time.Experiments on MovieLens dataset show that the accuracy of matrix decomposition model based on FTRL is only 2.3%lower than that based on SDG,but the training time is 10 times shorter.In order to alleviate the cold start problem of recommendation system and improve the accuracy of recommendation results,a new hybrid recommendation model based on deep learning is proposed in this paper.The model calculates the user's preference for items based on historical behavior and attribute information,that is,the user's rating on the merchant or the click probability of the merchandise.Historical behavior includes user rating data and click history.Attribute information refers to user's age and business type.After obtaining the implicit factor vectors of users and items,the model does not use linear models such as inner product or cosine similarity to calculate preferences,but takes full account of the feature crossover between users and items,thus improving the accuracy of prediction results.Experiments on MovieLens dataset show that the proposed model has advantages in recommendation accuracy and model complexity.Finally,we apply the matrix decomposition model based on FTRL and the proposed hybrid recommendation model to the design of business recommendation algorithm for green trip APP.Green travel recommendation algorithm is divided into two stages:recall stage and prediction stage.In the recall phase,a matrix decomposition model based on FTRL is used to select a small number of merchants that may be interested by users from all merchants and provide them to the ranking stage.In the sorting stage,the merchants selected in the recall stage are scored and sorted by the proposed hybrid recommendation model,and a list of merchant recommendations is generated for users.
Keywords/Search Tags:Recommendation Algorithm, Collaborative Filtering, Matrix Factorization, Hybrid Recommender Mode
PDF Full Text Request
Related items