Paper reading - Learning Centrality by Learning to Route (LCR)

Preface: my postgraduate research topic is complex network, so I need to read some thesises related to it.

Paper DOI: 10.1007/978-3-030-93409-5_21

1. Introduction

General: This paper proposed a method to use a fully connected deep learning to learn the route function in Route Betweeness Centrality (RBC).

• route function: four inputs and one output function. usually marked as R(s,t,u,v), among which s means source vertex, t means target vertex, R(s,t,u,v) is the probability that transmiting a package from s to t and this package pass through u to v edge.
• relationship between route function and RBC: by sumarizing the weighted ($T(s,t)$) route function result of certain point we can get it RBC value:
  $$RBC(v)=\sum _{u,v\in allvertex}T(s,t)\times R(s,t,u,v)$$

2.Main

The main structure of Learning Centrality by Learning to Route (LCR) is as above.

• $n^4\times 4d$: a table to store all permutations of vector { s , t , u , v } \{s,t,u,v\} {s,t,u,v}, total $n^4$ elements and each element is a 4-d vector.
• Routing Policy: a 4-d table to store all the results of routing function.
• Computing method depends on route method: there are different approaches to get RBC value, such as Load Centrality (Load), Shortest Path Betweenness Centrality (SPBC), closeness, degree, and eigenvector centralities on the various graphs.
• Eigenvector RBC Algorithm:
  
  • traffic matrix: shape is same as adjacency matrix, to denote the propotion of packages send from s to t in all packages.

3.Result

It learns with good precise.

4.Furture to do

1. to improve the complexity of differentiable centrality computation to utilize the full power of deep learning on graphs.
2. learn a node embedding function alongside the routing function to apply the proposed approach on arbitrary graphs.
3. use geometric embedding algorithms to produce node positions suitable for applying LRC or arbitrary graphs.
4. explainable artificial intelligence algorithms can be used to visualize the learned routing functions and understand the structural properties that facilitate target variables, such as social influence.

5. Reference

1. Bachar, L., Elyashar, A. and Puzis, R., 2021, November. Learning Centrality by Learning to Route. In International Conference on Complex Networks and Their Applications (pp. 247-259). Springer, Cham.