Real-time Graph Partition and Embedding of Large Network

abstract

graph embedding is a method to embed a high dimensional graph into a much lower dimensiona vector space while maximally preserving the structural information of the original network.

we address a real-time and distributed graph embedding algorithm (RTDGE) : distributively embedding a graph in a streaming fashion      contain : 1) a graph partition scheme     2)a dynamic negative sampling method     3)unsupervised global aggregation 

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

limitations of the existing method :

   1) can not perform real-time streaming data

   2) most of the model are centralized ,which means that they can not handle big data   

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

three steps of  this model : 1) graph partition    2) dynamic graph embedding   3) graph aggregation

contributions:

   1) divide edges into non-overlapping subgraphs

   2) distributed sugraphs adaptively update the embedded vectors with incoming edges

   3) contruct a distrubuted big-data real-time processing platform to evaluate our model

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

model :      

           :  directed link from vertex vi to vj ,    weight wij

    optimization :     (1)       

                   are the embedded vectors for vi and vj

                  s( ) is the predifined similarity function (this paper use the function defined in LINE  ,  first-order and second-order                         proximity)

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

problem formulation :

       1)  first-order similarity :    (2)    W : the total weight of edges

                            after mebedding :    (3)

                           d( ) is the KL distance    (1) can be written as :      (4)

                           plug (2) (3) into (4) and apply KL distance :    (5)  ???????

                           KL distance : https://www.cnblogs.com/ywl925/p/3554502.html

     2)  second-order similarity : 

                                      

                                       N(vi) : vi's neighborhood

                           0

                        (1) can be written as :   

                        

                                                             

 

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

graph parittion : 

             1) initially, all K subgraphs are open            will be closed, if its capacity is reached （maximal number of edges）

            2）an incomming edge e=(u,v) will be assigned to the subgraph who has the minimal weight among all subgraphs                            containing  vertex u or v

           3)  te is defined to balance the number of edges :    

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

dynamic graph embedding :

    for a vertex :

             1) positive samples :  vertices directly connected to it

             2) negative samples : without direct connections to it

the optimization problem become : 

                                

where  ：                             

                                   

       f(v)  the weight sum of the selected negative samples       

      α is the smoothing parameter  

未完 待续。。。