| Recommender systems have attracted a lot of recent attentions and are widely used in the Internet due to their great commercial value.To promote sales and improve user experience,web sites such as A-mazon,EBay,and Taobao rely on recommender systems to analyze users' purchasing records,predict users' preference(ratings)to different items,and recommend those with the highest ratings to attract users.Despite its efficiency in discovering and quantifying the interactions between users and items,matrix completion in recommender systems suffers from the problems of low rating density and scalability.To conquer the above challenges as well as to provide quick and high quality recommendation,we propose oLSHpMatrix,a novel recommender scheme which concurrently exploits location-sensitive hashing(LSH)and matrix completion.In this paper,we primarily discuss the following points:1.Taking advantage of the good property of LSH,oLSH adopts an LSH hash table to reorder users in the system with similar users buffered in close positions.Without involing quadratic time complexity as that of the traditional memory-based approach,the LSH hash table can manage users to query of similar users can be performed on-the-fly.The key idea of oLSH is to hash the users using several hash functions,and for each function,the probability for the hashed values to fall into the same bucket is much higher for similar users which have close preferences than for users whose preferences are far apart.2.Based on the LSH hash table,oLSHpMatrix partitions the original user-item matrix into sub-matrices which contains similar users that impose higher impacts on each other.The submatrices formed can better capture the user interests and have high correlation thus low rank,and the later helps to more accurately reconstruct the sub-matrices.further increase the sampling ratio and reduce the scale of the sub-matrix,we also propose to prune the columns which are not much related to the target users.Compared to the original user-item matrix,interactions between users and items in the sub-matrix significantly increase,therefore,quicker and higher quality recommendation can be achieved.We have done extensive experiments using real-world data sets to compare oLSHpMatrix with the state of art recommender schemes.Our experimental results demonstrate that oLSHpMatrix can achieve much better recommendation performance with its use of partitioned matrix which is of much smaller scale and has larger rating ratio. |
