We hide the associations in one group and run the methods on the remaining associations, repeating three times to ensure that each group is hidden exactly once.
The parallell implementation was not strictly necessary, but made it much more feasible to run the methods on all the relatively large datasets.
We then ran the method of Braga et al. (2008) on the simulated sequences (comparing each time the initial sequence and one sequence with simulated inversions).
Even for newer methods such as ADMIXTURE or sNMF, finding the optimal value of K relies on running the methods for a range of values of K.
In order to measure the dependance of MowgliNNI on the precise costs used for each kind of event, we ran the method with the threshold t = 80 on G ML trees with costs varying up to 10%, 20%, then 50% w.r.t.
At each fold of the CV, a linear model was derived by running the proposed method on 2/3 of the data in the dataset, and then tested on the remaining 1/3 of the data.
Solutions obtained running the proposed method on simple synthetic datasets allowed to identify a recurrent pattern, which has been found in more complex, synthetic and real, datasets.
The results of running the STL method on the same date range, but using only 90 days of baseline data for each scenario, are shown in final column of Table 1.
Running the proposed method on the GPU requires ~40 h to complete as compared to ~57 h for completion using F astE pistasis on 48 AMD Opteron 6172 2.1 GHz CPU cores (Sunnyvale, CA, USA) with ATLAS BLAS/LAPACK version 3.2.1 (Whaley and Petitet, 2005).
We randomly split the dataset into 23 training and 24 test samples 50 times and ran all the methods on each partition.
