Fig. 2
Mean clustering performance. The mean performances (with ± s.d. error bars) are laid on top of the individual cluster performances. The clustering methods are organized in ascending order for each performance measure. Under the UFL legend, “OFF” indicates that only L2-normalized features were used for clustering. This means that no feature extraction was performed. a Mean NMI across all datasets. b Mean ACC across all datasets. The MANOVA revealed that there was a statistically significant difference in applying ICA BSS after the matrix factorizations (Pillais’ Trace = 0.058, F(1,166) = 4.05, p < 0.05). The post hoc pairwise t-tests show that ICA BSS provides higher mean performance across all datasets (NMI: µICA-ON = 0.783 ± 0.174 and µICA-OFF = 0.716 ± 0.20, p < 0.05; ACC: µICA-ON = 0.745 ± 0.133 and µICA-OFF = 0.653 ± 0.165, p < 0.001). MANOVA also showed there was significant difference in clustering method (Pillais’ Trace = 0.54, F(6,166) = 8.14, p < 0.001). GNMF, SPC-RW, SPC-SYM, and ICA-SYM had the higher mean performance compared to all other clustering methods (all comparisons p < 0.001), but there was no statistical difference among these four methods. The UFL processing component was significant at the less stringent p = 0.1 level (Pillais’ Trace = 0.063, F(2,166) = 2.14, p < 0.1). Post-hoc pairwise t-tests at the p < 0.05 level show that the NMI performance is higher with RICA (µUFL-RICA = 0.774 ± 0.137) and SFT (µUFL-SFT = 0.786 ± 0.144) feature learning in comparison to the absence of feature learning (µUFL-OFF = 0.706 ± 0.238). All post hoc pairwise t-tests were corrected using the False Discovery Rate