Download presentation
Presentation is loading. Please wait.
Published byClara Richard Modified over 8 years ago
1
Energy Models for Graph Clustering Bo-Young Kim Applied Algorithm Lab, KAIST
2
Introduction Graph clusters : groups of densely connected nodes A common structure shared by many real-world system Decomposed into subsystems : strong intra-subsystem interaction, weak inter-subsystem interaction. e.g. groups of friends, cohesive modules in software systems. Graph layout : nodes → positions in low-dimensional Euclidean space Goal: Introduce and evaluate measures(=energy models) that quantify how well a given graph layout reflects the graph clusters Node-repulsion LinLog Edge-repulsion LinLog
3
Introduction Interpretability vs. Readability The most energy based graph layout technique : produce easily readable box-and-line visualization of graphs e.g. Goal of Fruchterman and Reingold model(1991) Interpretable layout : positions of Euclidean distances of nodes reflect certain properties of the graph The density of subgraphs (This case) Graph-theoretic distances of nodes Direction of edges in directed graphs
4
Basic Definitions A graph G=(V,E) Assume that connected, undirected, no edge weight A d-dimensional layout of the graph G a vector of node positions The distance of u and v in p :
5
Clustering Criteria A subgraph with many internal edges and few edges to the remaining graph a graph cluster Coupling measures: a smaller coupling a better clustering The cut The node-normalized cut The edge-normalized cut Shi and Malik’s normalized cut Expansion Conductance Newman’s modularity
6
The Cut The cut for two disjoint sets of nodes V 1 and V 2 Expected cut between V 1 and V 2 : Biased towards bipartitions : smaller with | V 1 | ≪ | V 2 | than | V 1 | = | V 2 |
7
The Node-Normalized Cut The node-normalized cut (=density of the cut, the ratio of the cut) Expected value: Still biased towards bipartitions (measure of size : # edge) smaller with deg(V 1 ) ≪ deg(V 2 ) than deg(V 1 ) = deg(V 2 ) ( ∵ Expected value: )
8
The Edge-Normalized Cut The edge-normalized cut Expected value:
9
Other Clustering Criteria Shi and Malik’s normalized cut biased toward small clusters when is not fixed.
10
Other Clustering Criteria The expansion The conductance Bias The expansion : biased towards similarly-sized clusters The conductance : similar bias (cluster size measure : total deg)
11
Other Clustering Criteria Newman’s Modularity Two sets case If V 1 ∪ V 2 =V, maximizing Q’(V 1,V 2 ) = minimizing
12
The LinLog Energy Models The node-repulsion LinLog energy of a layout p The edge-repulsion LinLog energy of a layout p Symmetry : Edges cause both attraction and repulsion
13
Separation of Clusters The arithmetic mean and the geometric mean G=(V,E), F ⊆ V (2), a layout p Degree-weighted geometric mean :
14
Separation of Clusters (Proof) Let the layout p 0 be a solution of the minimization problem: Let ⇔ ⇔ ∙∙∙(1)
15
Separation of Clusters ⇔ ⇔ ∙∙∙(2)
16
Separation of Clusters (Proof) Similar to the Theorem 1.
17
Interpretable Distance between Clusters (Proof) p 0 : a layout in P with minimum node-repulsion LinLog energy d 0 : the distance of V 1 and V 2 in p 0. The distance between all node in V 1 (resp. V 2 ) = 0 d 0 : a minimum of
18
Interpretable Distance between Clusters (Proof) Similar to the Theorem 3.
19
Example 1. Airline Routing
20
Example 2. World Trade
21
Example 3. Dictionary
22
Thank you
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.