Improving clustering performance using independent component analysis and unsupervised feature learning

Table 2 Cluster processing pipeline parameters used in experiments

Component^a	Feature extraction			Graph construction	Graph embedding			Clustering
Component^a	PCA	ICA BSS	UFL	kNN graph	Spectral clustering^b	GNMF	NMF	K-means
Hyper-param. 1	# PC’s = # classes	# IC’s = # classes	Method: RICA & SFT	k: 5	L_sym & L_rw	Max iters. = 100	Max iters. = 100	200 repetitions
Hyper-param. 2	–	–	# feats. = 256	σ = mean of kth nearest neighbors	–	α = 100	–	–

PC: principal component. IC: independent component. PCA: The number of features is set to the number of classes. ICA BSS: The number of source signals is set to the number of classes. UFL: The number of features is 256 for both RICA and SFT, which was determined by the source code examples from [27, 28]. kNN graph: The nearest neighbor value k is set to 5, as in Cai et al. [7]. The σ scaling factor is set to the mean of value of all of the kth nearest neighbors from the similarity matrix [19]. On large datasets, a smaller set of observations from the data is used to calculate σ, using the method of estimating population proportions [51]. Spectral Clustering: Two types of normalized graph Laplacians were used, the symmetric Laplacian and random walk Laplacian [19]. GNMF and NMF: The default parameters as provided in the code available from Cai et al. [7] were used. K-means algorithm: 200 repetitions of the algorithm were used in order to obtain stable clusters
^a Some processing components have more than one hyper-parameter. Thus, a second hyper-parameter is provided where needed
^b Spectral clustering also refers to the use of eigendecomposition for the Laplacian matrix