Distance-aware DAG Embedding for Proximity Search on Heterogeneous Graphs

challenge: 

  1) how to efficiently generate DAGs 

   2) how to effectly learn DAG embedding for proximity search

propose : D2AGE (distance-aware DAG Embedding)

source code :  http://github.com/shuaiOKshuai

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

Proximity search on hetegeneous graphs aims to measure the proximity between two nodes

previous work : measure the proximity by paths connecting two nodes

                 weakness : linear sequences have limited  expressiveness  for the complex connections

propose a model : D2AGE

                                 learning a representation for the DAGs to encode the proximity between two nodes

Semantic proximity search : given a node a, rank other nodes according to some semantic relations

                                             Application : link prediction      circle suggestion   ...

previous work :   measure the semantic proximity by paths connecting two nodes

this paper :  looking for a richer structure than paths   →   DGs(directed graphs)

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

DAGs for proximity search:

     1) distance awareness in DAGs :   RNNs    

     2) the efficiency in online generating DAGs

                                   - BFS   low efficiency

  We propose D2AGE  : 

  1)  solve distance-aware : assign different weights to different predecessors

  2) DAG generation

               -1 sample paths as an approximation for the whole graph

               -2 feed the samples to a distance-aware DAG generation mechanism to output DAGs and node distance

               -3 develop D2AG-LSTM to recursively apply the distance discounts from multiple predecessors

               -4 given the training supervision      devise an end-to-end solution     

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

D2AGE

    training object : learn

                                        1) the parameter              2) a proximity embedding model         to embed