Research On Hybrid Recommendation Algorithm Based On Average Preference Weight

In the process of social comprehensive informationlization,the exponential growth of Internet scale has caused redundant information in the network,resulting in insufficient information utilization and low efficiency.In order to use the Internet information more richly and effectively,the recommendation system is proposed.The key of the recommendation system is the preference prediction,which processes and analyzes the user information with historical data of users' behaviors to predict users' behaviors.When users' needs are unclear,the recommendation system is greatly significant.Studying the results of the existing recommended algorithms,we propose several new methods and conducts comparative discussions.The specific work is as follows:First,we review the research background of the personalized recommendation system and discusses the collaborative filtering algorithm commonly used for the personalized recommendation system.We also compare and analyze the commonalities and characteristics of several collaborative filtering algorithms to pave the way for the newly proposed algorithm.Second,we propose a kNN-based Average Preference Weight(APW)model for user preference prediction.First of all,the core of the model is the confidence level of the specified items,which is defined as the complement of the relative score error between items(subtract this error from 1.0).The confidence between any two users is determined by the item evaluated by both sides and the score.When they have the same score for an item,the confidence between the users based on the item is 1.Next,there are often more than one items jointly evaluated by two users.Therefore,this paper takes the confidence expectation of all co-evaluation items as the true preference weight between two users,namely APW.Then,under the kNN framework,this paper focuses on the APW between the target user and any of its neighbors.It has directionality,that is,the weight of the neighbor to the target user rather than the other way around.Finally,because the APW focuses on the weight between the target user and its neighbors,it can be combined with all neighborhood-resolving technologies flexibly and naturally.This paper mainly discusses six major techniques based on the cosine,Pearson and Euclidean distance similarity.Third,the paper proposes an Item Gravity Recommendation(IGR)based on kNN.Analogous to the formula for "gravitational force",we put forward three new item quality definitions.When an item is rated by more users,the quality of itself will be greater.Two classic distance formulas have been introduced in this study,namely Euclidean distance and Manhattan distance.The gravitation between the two items can be obtained based on the quality and distance of these two items.The larger the value,the more similar the two items are,and vice versa.IGR gives the real physical meaning of the item similarity measure with better readability and interpretability.Finally,this paper uses the rating data of 943 users on 1,682 films from the MovieLens public data set for comparison through different weighting algorithms.With the mean absolute error(MAE)and root mean squared error(RMSE)as indicators,the paper compares the original recommended algorithm with the improved six algorithms to verify the effectiveness of algorithms and test the performance of the IGR algorithm.Experimental results show that some of kNN-based APW algorithms have better effect than the original algorithm.The algorithm with the integration of the average preference weight can improve the existing algorithm to some extent.