Research On Interaction-based Collaborative Filtering Algorithm

The increasing popularity of the Internet has promoted the explosive growth of E-commerce.However,the rapid expansion of the network data makes it difficult to obtain useful information and services.How to solve the problem of information overload and help users to find the resources of interest has become a hot spot of current research.One of the most effective ways to solve the issue is the recommender system(RS).Collaborative filtering(CF)is the core and most widely used algorithms in RS.The base idea of CF is to find neighbors among users or items through the users' rating data of items,so that items can be recommended based on the neighbors' preference.However,most of the collaborative filtering algorithms still follow the traditional similarity measures in the obtaining of similar neighbors,and rarely consider the interaction between the user and the RS during the process of recommendation,which limits the accuracy and efficiency greatly.Therefore,it is of great practical significance to study the use of novel similarity measures to recommend in the user-RS interactive scenario.In this paper,we present a collaborative filtering algorithm based on user-recommender interactive scenario,which provides users with personalized recommendation of interest.First,we propose a similarity measure named Triangle,and combine it with the traditional Jaccard and Cosine similarity,then we define two JCT similarity measures named JCT_A and JCT_M.Jaccard is defined as the size of the intersection divided by the size of the union of the sample sets,and is used to measure the similarity between sample sets.The higher value of Jaccard indicates the more common items that the neighbors have rated,and the more reliable the similarity is.Cosine shows the rating preference of neighbors for different items by the cosine of two rating vectors.The higher value of Cosine means less angel of two rating vectors,which indicates the more consistent the ratings preferences of users are.Triangle similarity is sensitive to the absolute value of the ratings,and the higher value of Triangle means the smaller gap between two ratings.The JCT_A adds up the jaccard,cosine and triangle similarities,while the JCT_M multiplies them.Second,we design and realize a batch feedback user-recommender interactive scenario.When the users log in the RS randomly,they will browse all the interesting items in the recommendation list.Then,users give feedback on their choices and ratings to the system.The RS will provide more accurate and diverse recommendations for user based on their feedback and history rating data.The scenario can efficiently recommend a number of items that are of interest to the users,and at the same time,meet the need that information providers want to recommend their resources to users as much as possible,which enables the system certain ability in mining the long tail items.Third,we evaluate the effectiveness of collaborative filtering algorithm based on user-RS interaction by two series of experiments on MovieLens 100K,MovieLens 1M,Each Movie and Dou Ban datasets.The first set of experiments compares the predicting ratings accuracy of seven similarity measures liked PIP?NHSM?JCT_A and JCT_M by MAE and RMSE.The second set of experiments contrasts the performances of Top N recommendations of single-feedback and batch feedback user-RS interactive scenarios,and then evaluates the performances of four similarity measures which contains Cosine?Pearson?JCT_A and JCT_M in batch feedback user-RS interactive scenario by recall?precision and coverage.The experiments of rating prediction show that JCT_M performs better than other similarity in MAE and RMSE on MovieLens 100K,MovieLens 1Mand Dou Ban datasets and JCT_A gains the least value of MAE and RMSE on Each Movie dataset.In Top N recommendation,the batch feedback user-RS interactive scenario obtains higher recall?precision and coverage than single one;JCT_A on Top N recommendation performs better than the other similarities.