Figure 2

Average bias μ, standard deviation of bias σ and the Pearson coefficient ρ against the data amount parameter η for the MovieLens, Amazon, Last.FM and Epinions data sets respectively.

Each data point is averaged over 20 independent experiments, i.e., for each level of data amount parameter η, we randomly divide the data for 20 times and calculate μ, σ and ρ of each time. Note that, there is only selecting information without ratings in the Last.FM data set. Thus, the COS and PC indices could not be performed in the Last.FM data set. As the data becomes more and more abundant, the stability of object similarity would relatively be better. However, many measurements still could not give stable evaluation of object similarity.