Improvement On Recommender Systems

With the rapid development of network and information technologies,more and more users want to find the information they need on the Internet,including actively and passively finding such information.However,data about users,commodities,transaction records,and social information on the Internet has grown explosively.The amount of information far exceeds our ability to process information.Information overload is becoming more and more serious.Now it takes more and more time and effort for users to find the target product or information.The recommendation system is an effective solution to the problem of information overload and the research on the recommendation system has received more and more attention.Scholars and industrial researchers have done a lot of researches on the recommender system,including the accuracy,diversity,novelty,etc.of recommendation.In particular,the emergence of data competitions such as the Netflix Prize,Kaggle,Alibaba mobile recommendation algorithm contest,etc.,hasstrongly promoted the innovation and development of the recommendation algorithm.The recommended algorithms are divided into content-based recommendation algorithms,knowledge-based recommendation algorithms,demographic-based recommendation algorithms and collaborative filtering recommendation algorithms.Collaborative filtering algorithm is a popular research domainby mining collective behavioral datasuch as transactions and ratings that can reflect users' interests.Collaborative filtering algorithms can be further divided into neighbor-based collaborative filtering algorithms and model-based collaborative filtering algorithms.The model-based collaborative filtering algorithm is the main research branch of collaborative filtering algorithms.According to the differences in models,there are collaborative filtering algorithms based on the clustering model,graph model,matrix decomposition model,neural network,etc.As a kind of easily accessible context information,time information can improve the quality of the recommendations.The recommender system not only needs to recommend the user with items that are in accordance with his interest,but also needs to ensure that when the user's interest changes,the recommendation system can discover changes about the user's interests and make adjustments to models in time.Considering the time factor in the recommendation system is one of the important improvement directions for the recommendation system.Some of the recommendation algorithms use time data as the basis for selecting the training set.Some of the recommended algorithms use the time data as a new feature of the training set and put the time variable into the model.At present,collaborative filtering recommendation algorithms that consider the time factor are mainly classified into two categories:Categorical time-aware algorithms and continuous time-aware algorithm.The categorical time-aware algorithm regards time as a discrete variable with limited values and it is considered that the user has the same interest when the time variable is the same and thus the user's interest is periodic;the continuous time-aware algorithm regards the time as a continuous variable,and considers the user's interest is decaying over time.This paper presents a new method for processing time data in collaborative filtering recommendation algorithm based on singular value decomposition model.Changes in user status and item status can be simulated by adding time data to the model.At present,in the researches of temporal SVD-based model,it is general to establish a model for a single user/item at each time in order to exhibit the change trend during the entire historical period.However,from the point of application,the historical model isoften useless and only the latest model is useful and describes the latest status of the users/items and will affect the user's future behavior.Generally,only the latest model is used for the prediction of future user behavior.At the same time,since the vast data and thescarce computational resources in practical application scenarios,the algorithm needs to reduce the computational complexity and time complexity as much as possible.Therefore,this paper only models the users/items of the latest moment.Considering that the user's recent behavior can better reflect the latest status of the user/item,the long-term behavior contains less information about the latest status of the user/item,thispaperassigns different weights to the behaviors at different moments to describe the amount of information that these historical behaviors have at the moment.Ebbinghaus forgetting curve represents the law that people's hobbies or memories will gradually decay with time.This paper fits the information retained by the user's historical behavior at the latest moment by the forgotten curve.Netflix dataset is a dataset widely used by the researchers of collaborative filtering domain because of its large volume and authenticity.This paper verifies the performance of the proposed algorithm on the published Netflix dataset and compares it with the algorithms proposed by former researchers.It is proved that the proposed algorithm in this paper can further improve the accuracy withoutadding the computational complexity of the collaborative filtering recommendation algorithm.