A careful assessment of recommendation algorithms related to dimension reduction techniques

摘要

Due to lack of detailed tests under the uniform test framework for existing personalized recommendation algorithms, the best performances claimed in literatures are hard to credit. For this reason, this paper presents a comparative evaluation of eight collaborative filtering (CF) algorithms, mainly focusing on recommendation algorithms related to dimension reduction techniques, on two common popular datasets by using three quality metrics. The eight algorithms are the k-nearest neighbor (KNN) algorithm, three native dimension-reducing algorithms respectively based on singular value decomposition (SVD), non-negative matrix factorization (NMF) and weighted non-negative matrix factorization, and four hybrid algorithms respectively crossing principal component analysis: PCA and KNN, SVD and KNN, NMF and KNN, and PCA and a recursive rectangular clustering. There are some interesting findings in our experiments. First, dimension-reducing techniques can help dig out more valuable information from the rating data than the nearest-neighbor technique. Second, in comparison with four hybrid algorithms, three native algorithms only based on dimension-reducing techniques are able to better satisfy users’ actual needs. Third, dimension-reducing techniques with non-negativity constraints are more effective than not with non-negativity ones. Fourth, the decision on the optimum algorithm among eight algorithms is insensitive to the sparsity of dataset. Fifth, proper selections for appropriate values of parameters of algorithms are very often problem dependent. Sixth, native dimension-reducing algorithms can defeat these algorithms related to KNN in the processing time. These findings not only show that it is necessary to evaluate in detail these algorithms under the uniform test framework but also play an important role in the right navigation for further innovation research on the technology of the personalized recommendation.