Subspace Segmentation via Quadratic Programming (SSQP) solves the following problem $$\textrm{min}_Z \|X-XZ\|_F^2 + \lambda \|Z^\top Z\|_1 \textrm{ such that }diag(Z)=0,~Z\leq 0$$ where \(X\in\mathbf{R}^{p\times n}\) is a column-stacked data matrix. The computed \(Z^*\) is used as an affinity matrix for spectral clustering.
SSQP(data, k = 2, lambda = 1e-05, ...)an \((n\times p)\) matrix of row-stacked observations.
the number of clusters (default: 2).
regularization parameter (default: 1e-5).
extra parameters for the gradient descent algorithm including
maximum number of iterations (default: 100).
tolerance level to stop (default: 1e-7).
a named list of S3 class T4cluster containing
a length-\(n\) vector of class labels (from \(1:k\)).
name of the algorithm.
Wang S, Yuan X, Yao T, Yan S, Shen J (2011). “Efficient Subspace Segmentation via Quadratic Programming.” In Proceedings of the Twenty-Fifth AAAI Conference on Artificial Intelligence, AAAI'11, 519–524.
# \donttest{
## generate a toy example
set.seed(10)
tester = genLP(n=100, nl=2, np=1, iso.var=0.1)
data = tester$data
label = tester$class
## do PCA for data reduction
proj = base::eigen(stats::cov(data))$vectors[,1:2]
dat2 = data%*%proj
## run SSQP for k=3 with different lambda values
out1 = SSQP(data, k=3, lambda=1e-2)
out2 = SSQP(data, k=3, lambda=1)
out3 = SSQP(data, k=3, lambda=1e+2)
## extract label information
lab1 = out1$cluster
lab2 = out2$cluster
lab3 = out3$cluster
## visualize
opar <- par(no.readonly=TRUE)
par(mfrow=c(1,3))
plot(dat2, pch=19, cex=0.9, col=lab1, main="SSQP:lambda=1e-2")
plot(dat2, pch=19, cex=0.9, col=lab2, main="SSQP:lambda=1")
plot(dat2, pch=19, cex=0.9, col=lab3, main="SSQP:lambda=1e+2")
par(opar)
# }