PRISMA (version 0.2-1)

prismaNMF: Matrix Factorization Based on Replicate-Aware NMF

Description

Matrix factorization $A = B C$ with strictly positiv matrices $B, C$ which minimize the reconstruction error $\|A - B C\|$. This replicate-aware version of the non-negtive matrix factorization (NMF) is based on the alternating least squares approach and exploits the replicate information to speed up the calculation.

Usage

prismaNMF(prismaData, ncomp, time = 60, pca.init = TRUE, doNorm = TRUE, oldResult = NULL)

Arguments

prismaData

PRISMA data for which a NMF should be calculated.

ncomp

either an integer or prismaDimension object specifying the inner dimension of the matrix factorization.

time

seconds after which the calculation should end.

pca.init

should the $B$ matrix be initialized by a PCA.

doNorm

should the $B$ matrix normalized (i.e. all columns have the Euclidean length of 1).

oldResult

re-use results of a previous run, i.e. $B$ and $C$ are pre-initialized with the values of this previous matrix factorization object.

Value

• prismaNMF
  Matrix factorization object containing the $B$ and $C$ matrix.

References

Krueger, T., Gascon, H., Kraemer, N., Rieck, K. (2012) Learning Stateful Models for Network Honeypots 5th ACM Workshop on Artificial Intelligence and Security (AISEC 2012), accepted

R. Albright, J. Cox, D. Duling, A. Langville, and C. Meyer. (2006) Algorithms, initializations, and convergence for the nonnegative matrix factorization. Technical Report 81706, North Carolina State University

Examples

data(asap)
print(asap)
plot(asap)
asapDim = estimateDimension(asap)
print(asapDim)
plot(asapDim)
# let it run for 60 second
asapNMF = prismaNMF(asap, asapDim, time=60)
plot(asapNMF, minValue=0.2)
# ...and for another 60 seconds re-using the results from the previous run
asapNMF = prismaNMF(asap, asapDim, time=60, oldResult=asapNMF)
plot(asapNMF, minValue=0.2)
labels = getMatrixFactorizationLabels(asapNMF)