Comparison of different methods.

Results are obtained by 10-fold cross-validation experiments on MovieLens (5,000 users and 5,977 objects) with cosine similarity measure. * and *** denotes the method in the corresponding row is better than that in the corresponding column at the statistical significance level of 10−1 and 10−8 after the Bonferroni correction, respectively.+denotes the null-hypothesis cannot be rejected at the significance level of 10−1 after the Bonferroni correction. Restart probabilities for random walk approaches are set to 0.9. Parameters are set to ß = 10 for RWPL, λ = 0.04 for RWNN, δ = 0.17 for RWTF, ß = 10 for USPL, λ = 0.03 for USNN and δ = 0.20 for USTF to obtain the highest performance.Comparison of different methods.