# Table 5 List of features used in the prediction of a directed link between generic usersuand v, along with their description

Feature Description Rank
Location Binary attribute, whether u and v belong to the same city 14
Gender Binary attribute, whether u and v belong to the same gender 15
Age Absolute difference of ages 12
Library Cosine similarity between library vectors 5
Groups Cosine similarity between group membership vectors 7
Group size Size of the smallest group the two users have in common 6
Vocabulary Cosine similarity between sets of tags used 16
Contact list Cosine similarity of the vectors of social contacts 2
Outdegree Sum of the out degrees ($k out (u)+ k out (v)$) 11
Preferential attachment Product of the out degrees ($k out (u)⋅ k out (v)$) 13
Common neighbors Number of common neighbors, directed case ($CN(u,v)=| Γ out (u)∩ Γ in (v)|$) 4
Triangle overlap $C N ( u , v ) Γ out ( u )$ 1
Reciprocation Binary attribute, whether the inverse link (v,u) is already present 9
Resource allocation $∑ z ∈ ( Γ out ( u ) ∩ Γ in ( u ) ) ( 1 k out ( z ) )$ [37] 3
Local path Linear combination of common neighbors and common distance-2 neighbors ($CN+ϵ⋅C N 2$) [37] 10
Weighted flow $wf(u,v)=CN(u,v)+ ∑ x ∈ C N ( u , v ) min ( w ( u , x ) , w ( x , v ) ) C N ( u , v )$ 8
1. $Γ in / out (u)$ denotes the set of u’s in/out neighbors, $k out (u)=| Γ out (u)|$, and w(x,y) is the weight of the tie between x and y. The rank reported is the result of the Chi Squared attribute selection method applied to our test set; the bold font of the rank indicates that the corresponding feature has been selected for the restricted feature set.