PSpearman rank correlation. The Spearman rank correlation coeffi-
cient is another candidate for a similarity function [58]. For the
Spearman correlation, the items a user has rated are ranked
such that their highest-rated item is at rank 1 and lowerrated
items have higher ranks. Items with the same rating are
assigned the average rank for their position. The computation
is then the same as that of the Pearson correlation, except that
ranks are used in place of ratings.
Cosine similarity. This model is somewhat different than the previously
described approaches, as it is a vector-space approach
based on linear algebra rather than a statistical approach. Users
are represented as |I|-dimensional vectors and similarity is measured
by the cosine distance between two rating vectors. This
can be computed efficiently by taking their dot product and
dividing it by the product of their L2 (Euclidean) norms
by min{|Iu∩Iv|/50,1}.
Constrained Pearson correlation. Ringo solicited ratings from its
users on a 7-point scale, providing a rating guide that fixed 4
as a neutral (neither like nor dislike) value rz. With an absolute
reference, it is possible to correlate absolute like/dislike
rather than relative deviation (as the standard Pearson r does).
This led Shardanand and Maes [137] to propose the constrained
Pearson correlation: