The NNSR algorithm works with MPI which is the parallelization library employed for many supercomputer and/or PC cluster systems. See the document of your computer system about how to run MPI based parallel software.

By default, only the root rank (process) reads and writes input and/or output files, including log message files. Therefore, if the system employs distributed storage system, then you do not need to send input files to all the computation nodes. To specify all processes to output the log messages, use the -L 2 argument of INGOR. The log message file will be generated for all the processes. These are stored in the local computation host where the process runs.

Generally, an MPI program is executed through mpirun or mpiexec program:

$ mpirun -np procs INGOR_installed_path/ingor.X.X.X-mpi

If your system requires a job script, you may need to do this in your script file.

The following INGOR options (see INGOR Manual) are ingored by this algorithm.

-p (maximum parents)
: If you want to specify the maximum number of parents when sub network estimation, use the mp argument below. However, generally, specifying more than 1 for this parameter significantly decreases the accuracy of the estimated networks.

See NNSR for the internal implementation document.

Arguments

T=n
iterations=n: The number of subnetwork estimation. (default: n = 100000)
th=x: Cut off threshold of edge frequecies for the final structure. (default: x = 0.2)
m=n
module=n: Subnetwork size. (default: n = 1000)
args={key=value,...}: Arguments for the Greedy Hill-Climbing Algorithm.
ex=n: Exchange interval, that specifies to exchange the estimated networks stored in local processes in every n times of network estimation. n = -1 for automatic setting based on the size of MPI processes.
mp=n: The maximum number of parents in subnetwork estimation. Note: this is not the maximum number of the final network structure. (default: n = 1)
rw_th=x: Threshold for a network used in random walk.
score_v=n
: Specifies the score function verbose level. During the NNSR algorithm execution, this verbose level is set to -1 by default. The meaning of the value depends on the score function. But, basically, -1 means to suppress log messages as much as possible.
output=file_prefix: If specified, all the subnetworks estimated during the subnetwork estimation by the neighbor node sampling are output in files file_prefix.n where n represents the MPI rank (process) ID starting from 0. Thus, each process outputs subnetworks handled by them into a single file. A subnetwork is stored in the PaList format with node (variable) IDs. It can be read with "id=m" arguments with PaList file reader where. m needs to be the number of nodes (variables) in the dataset. Each subnetwork is separated by an empty line in the output files. The subnetworks of output files can be read by BSFilter (--bs). Note that the last line of the output file will be an empty line. Older INGOR versions may fail to handle this properly.

INGOR Manual

NNSR