Scalable Tempral Latent Space Inference for Link Prediction in Dynamic Social Networks

Linhong Zhu

Application :

              1) recommending friends in an social network e.g.

              2) suggesting potential hires in a professional network

===============================================================

temporal link prediction : given a sequence of graph snapshots  G1,G2......Gt,  predict the new links in t+1

the probability of a link  depends on the position of their space

==============================================================================

Our model        large moves are less likely

contribution :

          1) model the temporal link probability of node pairs

          2) address algorithmic issuses by developing the standard global BCGD algorithm

          3) a local BCGD algorithm , a global BCGD algorithm,  an incremental BCGD        faster

          4) conduct experiments over large real-life graphs, achieve good predictive quality in both sparse and dense networks

the change of the graph                   : latent space representation  for node set

:a row vector  the latent space representation of node u

================================================================================

assumption:

     1) temporal smoothness  :   change gradually

     2) network embedding   :   closer in the original space , closer in the temporal latent space

     3) latent homophily     :   closer  →  more likely to interact

=====================================================================================

problems:

     1) temporal latent space inference 

        find a latent space representation     that minimize the following  quadratic loss

            (1)

: penalizes node u for suddenly changing the location           

:   ignore all of the diagonal entries when computing

 

 

predict Gt+1 at time t+1

satify     

Z1,Z2.....Zt  →  Zt+q  →   Gt+1

====================================================================

standard BCGD algorithm

(1) can be decomposed into a linked part and a non-linked part

for each node u at time ,we focus on optimizing :

 

update Z with the following :