Research On Similarity Related Problems In Collaborative Filtering Algorithms

With the development of Internet and Information technology,obtaining information through Internet had become one of the main ways for people to obtain information.However,with the increasing abundance of Internet information resources,a phenomenon named "information overload" had emerged.It makes it more and more difficult to acquire information through Internet.Recommendation system is a good way to solve this problem.Collaborative Filtering technology is one of the most mature and widely used recommendation system algorithms,but there are still some problems,such as Cold Start,Data Sparsity,similarity update and so on.In order to solve the problems existing in the CF,this paper conducts the following research:1.In this paper,Adjusted Cosine(ACOS)similarity was adopted as the standard for similarity calculation,and weighted and bipolar ACOS similarity are introduced.The weighted ACOS similarity can solve the problem that ACOS similarity didn't consider the numbers of the common ratings by assigning different weights to users or items with different numbers of common ratings.By dividing users or items into two sets of like and unlike according to certain standards,the bipolar ACOS similarity discards some users or items with large difference in ratings,so the users' scores for the same item are more similar and the problem of large user item matrix can be alleviated to a certain extent.2.A kind of similarity called Scenario-based Adjusted Cosine similarity was proposed: the expert users and ordinary users had given ratings according to the application scenario to the items.And then calculate the two kinds of similarities,they are combined by different weights to get the final similarity formula.Since expert user ratings are available at startup,the problem of cold startup for new users was solved.3.In order to resolve the similarity update problem of the recommendation system,the basic,weighted and bipolar Adjusted Cosine similarity UCF and ICF dynamic algorithms were proposed,and the local similarity update was carried out by using the incremental learning method.In this paper,the complex similarity calculation formula were decomposed into several similarity factors,and the increments of these similarity factors were calculated separately,which were brought into the similarity calculation formula to obtain the locally updated similarity.This method only needs local similarity updates,so it reduces the computational workload and complexity of similarity updates.