Sparsity, scalability, and distribution in recommender systems

摘要: Recommender systems apply knowledge discovery techniques to the problem of making personalized product recommendations. These are achieving widespread success in e-commerce nowadays. The tremendous growth customers and products poses some key challenges for collaborative filtering (CF) based recommender systems. axe: producing high quality recommendations, performing many recommendations per second millions products, coverage presence data sparsity, meeting demands availability. New system technologies needed that can quickly produce even very large-scale problems. In this dissertation, we present our approaches address three such research challenges—sparsity, scalability, distribution. For first two challenges, perform experiments on real-world data. Our show improvements over basic CF-based algorithm. third challenge, provide a framework be extended implement distributed We tackle sparsity ways—by implementing model integrating content-based ratings into CF by applying alternate algorithmic sparsity. approach, semi-intelligent agents generate analyzing syntactic features item content. singular value decomposition (SVD) prediction algorithms item-based algorithms. results suggest both these capable addressing issue. different scalability issue—by using dimensionality reduction neighborhood formation incremental model-based techniques. In use SVD technique clustering technique. SVD-based technique, association rule-based methods have potential improving distribution issue We taxonomy applications relative relevance geographically proximate distant users items. frameworks an evaluation each framework-application pair.