Nikos Manouselis et al. 431
wm = [wm wm wm ]
weights — ∣∙ 1 , 2 , 3 J that are associated with the three criteria. It is assumed that the
wm
weights i (i =1,..., 3) that a business partner m gives to each criterion remain the same
(that is, the partner considers the same criteria important when evaluating different e-mar-
kets). The evaluations gmi(k) are referred to as the evaluations of the business partner m, and
wm
the weights i as the properties of partner m.
• The total utility function of the e-market k for the business partner m is therefore:
um(k)=∑i3=1wimgim(k), ∀m∈M(5)
The goal of the e-market recommendation algorithm is to provide to a particular business
partner a ∈ C that has not rated a particular e-market k ′ ∈ S (who we will refer to as the active
user), a ranking of the e-markets. In the light of the ‘neighborhood-based’ algorithms
discussed by Herlocker et al. (2002), we therefore design a recommendation algorithm that is
based on the following principle: it creates a ‘neighborhood’ of D ⊆ M business partners that
a
have similar priorities to the properties wi of the active user, and examines how they have
evaluated k ′ (therefore, k ′ ∈ K should hold in order for the recommendation algorithm to be
able to produce a prediction). That is, it bases its recommendation on the opinion of the
business partners that assign similar importance weights to the evaluation criteria. If we
assume that z ∈ '∙l' is the number of members in the neighborhood, the goal of the
recommendation algorithm is to predict the total utilityu (k ) according to the z utilities
u (k ) of this e-market for each d ∈ D business partner in the neighborhood.
For this purpose, the similarity of the active user a to each user m′ ∈ M (denoted as sima,m’)
can be calculated using one of the classical measures used in recommendation literature
(Herlocker et al., 2002). In particular, we calculate similarity as the distance between the
vectors of the weights of the active user a ( — = [w1, w2, w3 ] ) and each business partner
m′ m′ m′ m′
m (w_ - Iw1 , w2 , w3 J), using the Cosine metric:
∑ ( w × O
sima m = =
, 7∑ (Wia)2 ×√∑3=,(wm')2 (4)
After the similarity of each business partner m′ with the active user a is calculated, the
neighborhood D of business partners from which the prediction of ua(k’) can be produced is
either formulated by selecting only partners with similarity over a pre-defined threshold (called
Correlation Weight Threshold) or from a pre-defined maximum number of users (Max
Neighbors Number). Both options can be considered in our algorithm.