Canopy Based Efficient Clustering for Collaborative Filtering

Abstract

Recommender systems are intellectual systems of information filtering system that makes relevant suggestions for customers. Main purpose is to generate only relevant information for customers to improve their experience. Recommendations systems mostly operate in challenging environment. A Recommender system which gives fast, accurate recommendations and good experience to users is more attractive and develop more interests in users. However recommender systems experience the ill effects of the ’sparsity’ and ’new user’ issues. It is frequently expected that information sparsity results in significant amount of corelated items or none between two users, bringing about inconsistent or inaccessible likeness data, and further bringing about helpless recommendation quality and decreased accuracy of predictions. High quality predictions are dependent upon how well the recommender system address its challenges. Thus there is this need to develop a model which should have improved performance under relatively highly sparse data. Clustering enhances the performance of recommender systems to counter the sparsity problems. Canopy Clustering (CC) is a simple but effective process for grouping objects into clusters. firstly, This thesis work proposes Density based Canopy clustering to enhance the performance of the traditional clustering such as KMeans algorithm by acquiring the value K and the initial clustering centroids, taking care of the two troublesome issues: the initial clustering centroid and the value K. The User-Item Category Preferred Ratio is also calculated to generate the UICPR matrix to cluster the data and as a result it helps to achieve better recommendations. Secondly, density based Canopy Clustering is also used on the user-movie rating matrix to increase the efficiency of k-means algorithm, Fusion of UBCF and IBCF on clustered data will further predicts the rating of items. Combining both these methods will help to achieve more accurate results. In this research Confidence based weighted fusion method (CBWF) is also presented to merge the rating predictions for hybrid collaborative filtering. Finally predictive accuracy and decision support metrics are used to compare the results on real world dataset i.e., MovieLens using RMSE, precision and recall.