Improving clustering performance using independent component analysis and unsupervised feature learning

Table 4 Comparison of clustering performance across different datasets and clustering techniques

Baseline	Performance of applied processing on different datasets (NMI, ACC)
Baseline	COIL20	COIL100	CMU-PIE	USPS	MNIST	REUTERS-10K
Method
K-means	0.735, 0.597	0.822, 0.615	0.532, 0.239	0.659, 0.694	0.527, 0.553	0.356, 0.541
Deep learning
AE+K-means (2016)					–, 0.818	–, 0.666
NMF-D (2014)	0.692, –	0.719, –	0.920, 810	0.287, 0.382	0.152, 0.75
TSC-D (2016)	0.928, 0.899				0.651, 0.692
DEN (2014)	0.870, 0.725
DBC (2017)	0.895, 0.793	0.905, 0.775		0.724, 0.743	0.917, 0.964
IEC (2016)		0.787, 0.546		0.641, 0.767	0.542, 0.609
AEC (2013)				0.651, 0.715	0.669, 0.760
DCN (2016)					0.810, 0.830
DEC (2016)			0.924, 0.801	0.586, 0.619	–, 0.818	–, 0.722
DCEC (2017)				0.826, 0.790	0.885, 0.890
DEPICT (2017)			0.974, 0.883	0.927, 0.964	0.917, 0.965
JULE-SF (2016)	1.000,–	0.978, –	0.984, 0.980	0.858, 0.922	0.906, 0.959
JULE-RC (2016)	1.000, –	0.985, –	1.000, 1.000	0.913, 0.950	0.913, 0.964
VaDE (2016)					–, 0.945	–, 0.798
IMSAT (2017)					–, 0.984	–, 0.719
SpectralNet (2018)					0.924, 0.971
Non-deep learning
AC-GDL (2012)	0.865, –	0.797, –	0.934, 0.842	0.824, 0.867	0.017, 0.113
AC-PIC (2013)	0.855, –	0.840, –	0.902, 0.797	0.840, 0.855	0.017, 0.015
SEC (2011)				0.511, 0.544	0.779, 0.804
LDMGI (2010)				0.563, 0.580	0.802, 0.842
Ours^a	0.965, 0.93	0.962, 0.897	0.986, 0.937	0.868, 0.926	0.824, 0.882	0.460, 0.714

^a The maximum clustering performance obtained by the processing components proposed in this study is provided for each dataset. When available, both NMI and ACC are presented in each cell, where the first value is the NMI. If the cell is blank then the clustering method was not used on the dataset. Italic font within a row indicates the maximum performance obtained for a dataset. The full names and references of the compared methods are: Deep Embedding Network (DEN) [11], Discriminatively Boosted Clustering (DBC) [55], Infinite Ensemble Clustering (IEC) [56], Autoencoder-based Clustering (AEC) [10], Deep Embedded Clustering (DEC) [8], Deep Clustering Network (DCN) [13], Deep Convolutional Embedded Clustering (DCEC) [57], Deep Embedded Regularized Clustering (DEPICT) [14], Variational Deep Embedding (VaDE) [15], autoencoder with K-means clustering (AE + K-means) [8], Information Maximizing Self-Augmented Training (IMSAT) [58], NMF with Deep learning model (NMF-D) [59], Joint Unsupervised Learning (JULE) “-SF” and “-RC” [16], Task-specific Deep Architecture for Clustering (TSC-D) [60] Graph Degree Linkage-based Agglomerative Clustering (AC-GDL) [61] and Agglomerative Clustering via Path Integral (AC-PIC) [62], Spectral Embedded Clustering (SEC) [63], and Local Discriminant Models and Global Integration (LDMGI) [52]