intelligent database systems lab n.y.u.s.t. i. m. batch kernel som and related laplacian methods for...
TRANSCRIPT
Intelligent Database Systems Lab
N.Y.U.S.T.I. M.
Batch kernel SOM and related Laplacian methods for social network analysis
Presenter : Lin, Shu-Han
Authors : Romain Boulet, Bertrand Jouve, Fabrice Rossi, Nathalie Villa
Neurocomputing 71 (2008)
˜
Intelligent Database Systems Lab
N.Y.U.S.T.I. M.
2
Outline
Motivation
Objective
Methodology
Experiments
Conclusion
Personal Comments
Intelligent Database Systems Lab
N.Y.U.S.T.I. M.Motivation
Peasants of French medieval society about 90% of the population, but their community are anonymous related to a master, so
To extract a community structure (degree distribution, the number of components, the density, etc.)
To provide a organization of these small homogeneous social group
Could help historians to have a synthetic view of the social organization of the peasant communities during the Middle Ages.
3Fig. social network
Intelligent Database Systems Lab
N.Y.U.S.T.I. M.
4
Objectives
To explore the structure of a medieval social network modeled trough a weighted graph.
1. Defines perfect communities and uses spectral analysis of the Laplacian to identify them.
2. Implements a batch kernel SOM which builds less perfect communities and maps them.
Results are compared and confronted to prior historical knowledge.
Fig. perfect communities Fig. Final self-organizing map (7 * 7 square grid)
Intelligent Database Systems Lab
N.Y.U.S.T.I. M.Methodology – perfect community
5
Laplacian method
Spectral properties of the Laplacian (to find the perfect community)
Intelligent Database Systems Lab
N.Y.U.S.T.I. M.Methodology – perfect community (Cont.)
Clustering through the search of perfect communities
The perfect community is a subgraph, which all its vertices are
pairwise linked by an edge Has exactly the same neighbors outside the community
The rich-club occurs when the vertices with highest degree from a dense subgraph with a small diameter.
The central vertices is a set of vertex which connect the whole graph.
6Fig. perfect communities (circles), the rich-club (rectangle) and central vertices (squares).
Intelligent Database Systems Lab
N.Y.U.S.T.I. M.Methodology – batch kernel SOM
7
(Dis)Similarity measure : Almost perfect communities to graph cuts
Diffusion kernel : define a kernel that maps the vertices in a high-dimensional space
Intelligent Database Systems Lab
N.Y.U.S.T.I. M.Experiments
(A) Simple representation of the graph by Tulip
Two persons are linked together if: they appear in a same contract,
they appear in two different contracts which differ from less than 15 years and on which they are related to the same lord or to the same notary.
9Fig. Representation of the medieval social network with force directed algorithm. (615 vertices and 4193 edges)
Fig. Cumulative degree distribution (solid) of the weighted graph
Intelligent Database Systems Lab
N.Y.U.S.T.I. M.Experiments (Cont.)
(B) Clustering the medieval graph into perfect communities and rich-club
10
Fig. green level of the disk encodes the mean date of the contracts in which the members of the community are involved
(from black, 1260, to white, 1340).
Fig. density of the induced subgraph as a function of the number of highest degree vertices (log scale)
Fig. number of components of thesubgraph of perfect community and rich-club
as a function of the number of vertices with high betweenness measure added
Intelligent Database Systems Lab
N.Y.U.S.T.I. M.Experiments (Cont.)
11
Fig. Graph of the perfect communities by geographical locations (yellow: Flaugnac, blue: Saint-Julien, green: Pern, pink: Cornus, red: Ganic and orange: Divilhac).
Intelligent Database Systems Lab
N.Y.U.S.T.I. M.Experiments (Cont.)
12
(C) Mapping the medieval graph with the SOM
Fig. Final self-organizing map (7 * 7 square grid).
Self-organizing map of the main cluster.
Fig. Mean date for each cluster from black, 1260, to white, 1340.
Intelligent Database Systems Lab
N.Y.U.S.T.I. M.
13
Conclusions
The two approach can both provide elements to help the historians to understand the organization of the medieval society.
The two approach have distinct advantages and weaknesses. Kernel SOM can provides a notion of proximity, organization and
distance between the communities.
Kernel SOM organize all the vertices of the graph (not only the vertices that belong to a perfect community).
Perfect community approach is more reliable for local interpretations.
The definition of a perfect community is restrictive.