1/23 Graph Colouring through Clustering István Juhos 18 Aug, 2009 University of Szeged Hungary U niversi ty of Edinburgh István Juhos 2/23 Outline ● Problem definition ● Consider the solved problem ● Get an idea to create a similarity matrix ● Approximate the similarity matrix using available structures ● Define clustering algorithm 18 Aug, 2009 U niversity of Edinburgh István Juhos Graph minimal vertex colouring 18 Aug, 2009 U niversity of Edinburgh 3/23 István Juhos A colour algorithm example, a sequential greedy approach 4/23 Take an order of the nodes, assign the first available colour 165432 4-colouring 123456 3-colouring (minimum) 18 Aug, 2009 U niversity of Edinburgh István Juhos Questions 5/23 1) Can a graph be coloured with k number of colours? 2) What is the minimum k called chromatic number 3) What are the k-colourings, ? -colourings? 4) Can we construct a minimum colouring? (this talk) k-colouring is NP-complete. Try to approximate. 18 Aug, 2009 U niversity of Edinburgh István Juhos Graph descriptions 6/23 Usually only V and E are given without drawing by sets by adjacency matrix (dot=0) 18 Aug, 2009 U niversity of Edinburgh István Juhos Graph drawing 7/23 no edge crossing Colouring/clustering in a vector space 18 Aug, 2009 attractive forces: colours repulsive forces: edges How can we define forces? no colours initially use edges (adjacency matrix) U niversity of Edinburgh István Juhos 18 Aug, 2009 U niversity of Edinburgh István Juhos 9/23 Colouring matrices 3-colouring matrix (a solution) 18 Aug, 2009 U niversity of Edinburgh István Juhos 10/23 All optimal-colourings No symmetric colourings 18 Aug, 2009 U niversity of Edinburgh István Juhos 11/23 Sum optimal-colouring matrices 1/2 SUM 18 Aug, 2009 U niversity of Edinburgh István Juhos 12/23 Sum optimal-colouring matrices 2/2 SUM {1,3} => 3 {1,5} => 2 {3,5} => 2 {4,6} => 2 {5,2} => 1 {2,4} => 1 blue {1,4} => 0 red red 4th colour red green blue 18 Aug, 2009 blue red U niversity of Edinburgh red István Juhos 13/23 Sum matrix defines similarity SUM Large values attractive forces Small values repulsive forces Similarity matrix for clustering We do not have any optimal colouring matrices! But we can approximate the sum matrix using only the adjacency matrix (see later). 18 Aug, 2009 U niversity of Edinburgh István Juhos Sum matrix decomposition (approx.) 18 Aug, 2009 U niversity of Edinburgh 14/23 István Juhos 15/23 Problems with the sum matrix Sum matrix can contain conflicting suggestions. But extreme values are usually significant. {1,3} => 3 {1,5} => 2 {3,5} => 2 {4,6} => 2 {5,2} => 1 {2,4} => 1 18 Aug, 2009 significant significant {1,4} => 0 conflicting U niversity of Edinburgh István Juhos 16/23 Zykov-tree (hierarchical clustering) Use extreme values and modify the problem step-by-step merge add edge SUM 18 Aug, 2009 U niversity of Edinburgh István Juhos Zykov-tree clustering using Lovász-theta 17/23 repeat until reaching complete graph 1. approximate the sum matrix (Lovász-theta SDP solution, see next) 2. find some extreme values 3. merge vertices and/or add edges (Zykov-step) How can we approximate the sum matrix? 18 Aug, 2009 U niversity of Edinburgh István Juhos Approximate the sum matrix 1/3 (colouring matrix properties) 18/23 A colouring matrix is positive semi-definite 3 dim Describes orthogonal vectors in 3 dim. 18 Aug, 2009 U niversity of Edinburgh István Juhos Approximate the sum matrix 2/3 (an Integer SDP) 19/23 2 dim Hard problem 18 Aug, 2009 U niversity of Edinburgh (Meurdesoif) István Juhos Approximate the sum matrix 2/3 (Relaxed Integer SDP) 20/23 (Lovász) Polinomial time transformed SUM 18 Aug, 2009 U niversity of Edinburgh István Juhos Zykov-tree clustering using Lovász-theta (again) 21/23 repeat until reaching complete graph 1. approximate the sum matrix (Lovász-theta SDP solution) 2. find some extreme values 3. merge vertices and/or add edges (Zykov-step) 18 Aug, 2009 U niversity of Edinburgh István Juhos 22/23 Some results 18 Aug, 2009 U niversity of Edinburgh István Juhos 23/23 Cheers 18 Aug, 2009 U niversity of Edinburgh István Juhos