Collaborative filtering system based on classification and extended K-means algorithm

Collaborative filtering (CF) is one of the most successful recommending techniques. With the tremendous growth in the number of users and items, however, the system encounters two key challenges, decreased recommending quality and increased response time. New technologies are urgently needed to deal with such large-scale problems. To address these issues, we suggest constructing the item category system based on the user-item rating matrix, calculating the similarity between items and classes, extracting the neighbor-class set, and predicting user scores based on such neighbor-sets. Because the dimension of the item classes is far smaller than the one of the items, the algorithm' computational speed is enormously enhanced. To mitigate the harmful effects on the system's predicting accuracy given by item-class based algorithm, the paper puts forward clustering after classification and extended K-means algorithm to construct the items' accurate category system. The experimental results indicate that classification and extended K-means algorithm have brought promising effects on the system, which ensure considerable predicting accuracy, while in the meantime, provide dramatically better performance than traditional item-based CF. So, the algorithm is a good choice for large-scale recommendation system.