Skip to main content

# Table 1 Comparison of running times of multi-way algorithms

Algorithm Number of Runtime (s) Runtime (s)
components n g =1000 n g =2000
CCA K=3 0.57±0.11 1.15±0.12
K=12 0.58±0.14 1.34±0.13
JIVE j V =1, i V =(1,1) 28.44±4.04 23.84±4.18
j V =4, i V =(4,4) 137.77±44.94 173.20±60.35
tPCA (2,2) 0.57±0.17 1.35±0.24
(2,6) 0.61±0.15 1.25±0.23
tWFOBI (2,2) 0.65±0.16 1.50±0.21
(2,6) 0.76±0.14 1.53±0.25
tWJADE (2,2) 0.66±0.19 1.44±0.24
(2,6) 1.23±0.21 2.71±0.28
PARAFAC R=6 22.37±2.53 37.32±4.51
R=12 48.11±2.98 100.83±7.92
iCLUSTER K=3 79.28±14.16 595.06±70.67
K=12 114.28±30.02 688.74±166.85
1. Seven multi-way algorithms in terms of the running times to infer components of variation (runtime) in the simulation model considered in Fig. 2. Estimates are medians and median absolute deviations over 100 Monte Carlo runs for when the signal-to-noise ratio is 1 (i.e. noise level = 3 in Fig. 2). The second column specifies the parameter values for the number of components used in each algorithm. The first rows for each method are as follows. For CCA, three sets of canonical vector pairs (K=3) are shown. For JIVE, the rank of joint variation (j V =1) and rank of individual variation (i V =1) for each data type are shown. For TPCA, TWFOBI and TWJADE, we inferred two components for both the data type and sample dimensions. For PARAFAC, the rank of decomposition was R=6 and for iCLUSTER the maximum number of clusters K was set to 3. For the second rows, the total number of components is exactly matched (12) for all methods. The running times are reported for two scenarios differing in the number of genes n g , as indicated, and were obtained on a Dell PowerEdge R830 with Intel Xeon E5-4660 v4 2.2GHz processors