Research On Collaborative Filtering Recommendation Algorithm Based On Weighted Similarity

With the rapid development of Internet technology,the amount of information on the Internet has exploded.Therefore,for each user,it is difficult to find the information they are really interested in in such a huge amount of data.As one of the effective methods to solve information overload,personalized recommendation technology has been widely studied and used.Among them,collaborative filtering recommendation method is the most widely used.The neighborhood-based collaborative filtering recommendation method has received extensive attention and research in academia and industry due to its simple design,good interpretability,and strong stability.The core idea is to select their neighbors by measuring the similarity between users(items),and then generate recommendations based on the neighbors' preferences,so the quality of the selected neighbors will determine the quality of the system's recommendation results.When using and processing users(or items)with the same score,the differences between users and between items will not be distinguished.However,in daily life,a user will like certain items more than other users,and some users will like certain items more than others.The traditional similarity measure does not consider the influence of these preferences in the calculation,resulting in inaccurate similarity calculation and low recommendation accuracy.In order to solve the above problems,this paper mainly does the following work:1.Aiming at the problem that the importance of users to items and the importance of items to users are not considered in the traditional similarity calculation,a weighted similarity measure is proposed to quantify the similarity between users and between items;by defining the concepts of users' core items and core users of items,to study the difference in the importance of different items to users and different users to items;then,propose the CUIS algorithm(Core-User-Item Solver)to iteratively update the system's core users and core items,as well as the importance weight coefficients of users to items and items to users;finally,the fast convergence and optimality of the CUIS algorithm are proved through a rigorous theoretical framework,and the experiments on real data sets also verify the CUIS algorithm for different similarity measures can effectively converge to the optimal solution.2.Aiming at the question of whether the proposed weighted similarity measure can improve the recommendation accuracy,the core users and core items and two importance weights obtained based on the CUIS algorithm are studied,and three recommendation methods are proposed for verification,and each method can be divided into two recommendation modes:user-based and item-based.The recommenders based on core users and core items recommends users or items that are similar to core users or core items.The k-Means clustering recommenders based on core users and core items first clusters users or items,narrows the search range,and then recommends.The weighted similarity-based PAF recommenders recommends the most popular items(or most loyal users)in the neighborhood.Experiments use the proposed weighted similarity measure and traditional similarity measure,respectively,to compare from two cases of binary discrete scoring and continuous scoring.Experimental results show that the proposed weighted similarity measure outperforms existing similarity measure methods and can improve the recommendation performance.