Although, as expected, the performance of every method declined with decreasing dataset size, the magnitude and timing of this decline varied strongly per method.
Stochastic error is the inevitable consequence of using finite datasets, and decreases as datasets grow in size.
As we can see in Fig. 4, in the case that the data volume of each node changes from 100 to 200, the metric of the time-dependent indicator decreases, the dataset completeness increases slightly, and the correctness indicator increases.
To decrease the dataset loading time, two approaches have been employed.
This is demonstrated in Figure 1, which shows that the number of single-copy marker gene homologs (recA, gyrB and rpoB) identified by BLAST when compared against a re-sampled GOS dataset decreases more slowly with a decreasing expect-value cutoff compared with less well-conserved genes.
Hence, the selectivity of the combination of pharmacophores against its dedicated NRs activity decreased for datasets with a large number of ligands (and consequently with a large number of pharmacophores included in the combination) or with a number of active ligands largely outnumbering the number of decoys.
This resulted in 30% decreases of each dataset.
The output is a reduced number of x-values, i.e., the number of columns in the training dataset decreases.
Nevertheless, when the overlap gets more penalized, the number of significant biclusters for the Yeast dataset decreases.
Since the sample size of the dataset decreases drastically with N, the reported P-value is somewhat unreliable for N = 4.
The optimal value of β is obtained by starting with a value of 1, evaluate the performance of EM on a held-out dataset, decrease the value of β, and check if the performance improves.
