Rank one SVD: un algorithm pour la visualisation d une matrice non négative

Transcription

1 Rank one SVD: un algorithm pour la visualisation d une matrice non négative L. Labiod and M. Nadif LIPADE - Universite ParisDescartes, France ECAIS 2013 November 7, 2013

2 Outline Outline 1 Data visualization and Co-clustering

3 Outline Outline 1 Data visualization and Co-clustering 2 One Dimensional embedding: Problem formulation

4 Outline Outline 1 Data visualization and Co-clustering 2 One Dimensional embedding: Problem formulation 3 Rank one SVD algorithm (R1SVD)

5 Outline Outline 1 Data visualization and Co-clustering 2 One Dimensional embedding: Problem formulation 3 Rank one SVD algorithm (R1SVD) 4 Experimental analysis

6 Outline Outline 1 Data visualization and Co-clustering 2 One Dimensional embedding: Problem formulation 3 Rank one SVD algorithm (R1SVD) 4 Experimental analysis 5 Conclusion and future work

7 Rank one SVD: un algorithm pour la visualisation d une matrice non négative L. Labiod and M. Nadif LIPADE - Universite ParisDescartes, France ECAIS 2013 November 7, 2013

8 Data visualization and Co-clustering Notation and definition Data set representation Data set can be viewed as an m n matrix A Each of the m rows represents an object (individual) Each of the n columns represents a feature (or attribute) Each entry a ij represents the element of the A in the intersection between row i and column j. Data examples Documents/words data Categorical data Social network - bipartite graph

9 Data visualization and Co-clustering Data visualization The visualization of A consists in an optimal permutation of rows and columns data, which involves a data reorganization revealing homogeneous blocks. Bertin has described the visualization procedure as simplifying without destroying and was convinced that simplification was no more than regrouping similar things. Spath considered such matrix permutation approaches to have a great advantage in contrast to the cluster algorithms, because no information of any kind is lost. Arabie and Hubert have referred to similar advantages calling such an approach a non-destructive data analysis, emphasizing the essential property that no transformation or reduction of the data itself takes place.

10 Data visualization and Co-clustering Data visualization Data visualization methods. Hierarchical clustering Bond energy algorithm (BEA, McCormick et al. (1972)) Genetic algorithm : Stress objective (Niermann (2005)) A Reordred A Figure: Initial Data - reordered data

11 Data visualization and Co-clustering Co-clustering Simultaneously partition data objects and features. Direct Clustering (Hartigan, 1975) Double kmeans (Tao Li, 2001) Spectral Co-clustering (Dhillon, 2001) Nonnegative matrix tri factorization: A USV t (Ding et al, 2006) Latent Block models (Govaert and Nadif,2003) Despite the advantages of co-clustering, all methods require the knowledge of the number of blocks. Balanced data: data2 Reordred data: co clustering result

12 Data visualization and Co-clustering Illustrative Example: 16 townships data Table: 16 Townships Data. Townships Characteristics A B C D E F G H I J K L M N O P High School Agricult Coop Rail station One Room School Veterinary No Doctor No Water Supply Police Station Land Reallocation Table: Reorganization of townships and characteristics. Characteristics H K B C D G L O M N J I A P F E High School Railway Station Police Station Agricult Coop Veterinary Land Reallocation One Room School No Doctor No Water Supply

13 Data visualization and Co-clustering Illustrative Example: reordred data Table: Characterization of each cluster of townships by a cluster of characteristics. class of townships {H, K} {B, C, D, G, L, O} {M, N, J, I, A, A, P, F, E} class of characteristics {High School, Railway Station,Police Station} {Agricult Coop, Veterinary,Land Reallocation} { One Room School, No Doctor, No Water Supply} Table: Reorganization of townships and characteristics. Characteristics H K B C D G L O M N J I A P F E High School Railway Station Police Station Agricult Coop Veterinary Land Reallocation One Room School No Doctor No Water Supply

14 One Dimensional embedding: Problem formulation Problem formulation R1SVD problem formulation Given an m by n nonnegative data matrix A The adjacency matrix of a bipartite graph is: [ ] 0 A B = A T, (1) 0 from which we define a stochastic data matrix as follow, [ ] [ S = D 1 0 Dr B = 1 A Dr 0 Dc 1 A T where D = 0 0 D c ], (2) where D r and D c, are diagonal matrices such that D r = diag(a½) and D c = diag(a T½).

15 One Dimensional embedding: Problem formulation Problem formulation Power method Since S is nonnegative and stochastic, The Perron-Frobenius theorem tel us that the first vector is also nonnegative and constant. The power method is the well known technique used to compute the =½ leading eigenvector of S. The power method consists in the following iterative process: π (t) = S (t) π (0) and π (t) = π(t) (3) π (t) The right eigenvector corresponding to the uniform distribution ( 1 m+n,..., 1 m+n,..., 1 m+n )T. The corresponding left eigenvector π represents the constant left eigenvector of S so that π T½=m + n. In the matrix notation we have π = Sπ and S½=½.

16 One Dimensional embedding: Problem formulation Problem formulation Power method π (t) = S (t) π (0) and π (t) = π(t) π (t) At first sight, this process might seem uninteresting since it eventually leads to a vector with all rows and columns coincide for any starting vector. The corresponding left eigenvector π =½represents the constant left eigenvector of S. However our practical experience shows that. If we stop the power method after a few iterations, the algorithm would have a potential application for data reordering. The key idea : early stopping for the power method (4)

17 Rank one SVD algorithm (R1SVD) R1SVD algorithm R1SVD : a mutual reinforcement principle [ ] u Now, let us consider π =, where u R v m + and v R n +. The upper part of π.i.e u is for documents weight and the lower part v is for words weights. Exploiting now the diagonal structure of S, then we can write [ ] [ ] [ ] { u 0 Dr = 1 A u u = D 1 v Dc 1 A T r Av (a) 0 v v = Dc 1 A T u (b) (5) This iterative process starts with arbitrary vector u 0 and repeatedly performs the updates of v and u by alternating between formulas (a) and (b) given in equation 5 until convergence.

18 Rank one SVD algorithm (R1SVD) R1SVD algorithm R1SVD algorithm R1SVD algorithm Input: data A R m n +, D r and D c Output: u,v Initialize: ũ = Dr 1 A½,,u = ũ ũ repeat ṽ (t+1) = Dc 1 A T u (t) v (t+1) = ṽ(t+1) ṽ (t+1) ũ (t+1) = D 1 r Av (t) u (t+1) = ũ(t+1) ũ (t+1) γ (t+1) u (t+1) u (t) + v (t+1) v (t) until stabilization of u, v, γ (t+1) γ (t) threshold

19 A Reordred A Reordred Sr Reordred Sc Sorted u Sorted v Rank one SVD: un algorithm pour la visualisation d une matrice non négative Rank one SVD algorithm (R1SVD) R1SVD algorithm Principal steps of the R1SVD algorithm Compute the first singular vectors u of matrix D 1 r D 1 c A T Early stopping of the R1SVD A and v of matrix Sorting the vectors u and v by descending (or ascending) order Reorganize the rows and columns data according to the sorted vectors Illustrative example Figure: Data1: Data1 reordered according u and v, reordered Sr = AA T according u, reordered Sc = A T A according v.

20 Rank one SVD: un algorithm pour la visualisation d une matrice non négative Experimental analysis Experimental analysis R1SVD : data visualization and co-clustering A Reordred A Reordred Sr Reordred Sc 4 5 Sorted u x Sorted v x Figure: Data1: Data1 reordered according u and v, reordered Sr = AAT according u, reordered Sc = AT A according v. A Reordred A Reordred Sr Reordred Sc 4 5 Sorted u x Sorted v x Figure: Data2: Data2 reordered according u and v,reordered Sr = AAT according u, reordered Sc = AT A according v.

21 Experimental analysis Experimental analysis Is this reordering meaningful?. In order to be able to answer this question we use confusion matrices to measure the clustering performance of the co-clustering result provided by our method. R1SVD : data co-clustering Table: Confusion Matrix evaluation on rows and columns data. data1 (rows) data2 (rows) data1 (columns) data2 (columns)

22 Conclusion and future work Conclusion R1SVD : conclusion We have presented an iterative matrix-vector multiplication procedure called rank-one SVD for data visualization. The procedure is to apply iteratively an appropriate stochastic adjacency data matrix associated to a bipartite graph, compute the first leading left singular vector associated to the eigenvalue λ 1. Stopping the algorithm after a few iterations involves a visualization of data matrix into homogeneous blocks. This approach appears therefore very interesting in co-clustering context.