Research On Recommendation Algorithm Based On Tripartite Graph And Time Effect

The development of information technology and the Internet has led to a drastic expansion of information.People have entered the era of information overload from the era of information scarcity.Information overload poses a great challenge to both users and information producers.Recommender system is an effective solution to the problem of information overload.It can analyze the user interest based on the historical behavior of users and automatically recommend items to usres.Recommender system typically uses user behavior data,tags,and contextual information to predict user ratings or recommend items for users.This paper mainly studies the influence of different information combinations on the quality of recommendation results.Besides,we propose a type probability algorithm(Type Probability),which considers both long-and short-term interests of users.First,this paper constructs a tripartite graph about ratings and tag information,and extracts ten algorithms based on different combinations of information.The first two algorithms are the traditional collaborative filtering algorithm,which only consider the relationship between users and items.The middle four algorithms take user or item as center,which use rating and the corresponding tag information to predict ratings or generate a recommendation list for users.The last four algorithms are tag-based recommendation algorithm,which only consider different types of tag information.This paper analyzes the time complexity of above ten algorithms and uses MAE(mean absolute error)and RMSE(root mean square error)to compare the accuracy of different algorithms in predicting ratings.For Top N recommendation,the precision?recall and coverage are used to evaluate the performance of ten algorithms.The experimental results on MovieLens 100K and MovieLens 1M show that the item tag-item-user is the best combination and linearly fused information can effectively improve the accuracy of rating prediction and the quality of Top N recommendation.Second,we simulate the running scenario,and further study the impact of time effect on recommender system.This paper proposes a new type probability algorithm that consider both short-and long-term interest of users.Then,we compare the hit ratio of three non-personalized algorithms and four personalized algorithms.The experimental on MovieLens 100K demonstrate that time context information is crucial for improving hit ratio.After considering the time effect,the hit rate of the current hot algorithm is higher than that of other algorithms,and the hit rate of the type probability algorithm proposed in this paper is higher than that of most of the algorithms at a specific time window.