A proof-of-concept library for motif analysis using MDL techniques. It contains the methods described in the paper Compression as a Fast Measure of Network Motif Relevance.
Each release on github comes with a compiled JAR file which you can run as a command-line program. Download the latest one here. See the section examples below, for how to use it.
The command line provides some basic results, but to do real experiments, you will probably want to call the code directlyfrom your own program. The simplest way is to include it as a Maven dependency through jitpack. Just include the following repository in your pom.xml file:
<repositories>
<repository>
<id>jitpack.io</id>
<url>https://jitpack.io</url>
</repository>
</repositories>
and the following dependency:
<dependency>
<groupId>com.github.pbloem</groupId>
<artifactId>motive</artifactId>
<version>v0.1.XXX</version> <!-- check http://jitpack.io/#pbloem/motive for the latest version -->
</dependency>
Check the jitpack link above for linking from gradle/sbt/leiningen projects, and to see what the latest release is.
Have a look at the classes Compare and CompareLarge for hints on how to set up a motif experiment from within java code. For command line usage, see the next section.
Display usage information:
java -jar motive.jar --help
Run the "synthetic" experiment: create random graphs with injected motifs.
java -jar motive.jar --type synth --synth.repeats 3 --synth.n 50 --synth.m 600 --synth.instances 0,5,10
Run the "fast" experiment (ER model and EdgeList model) on a particular dataset, finding motifs of size up to (and including) 10:
java -jar motive.jar --type fast --file data.txt --minsize 3 --maxsize 10 --samples 1000000 --maxmotifs 30
The "full" experiment includes the precise DS model as well. This is a bit slower.
java -jar motive.jar --type full --file data.txt --minsize 3 --maxsize 5 --samples 100000 --maxmotifs 30
The default data format is a text file with a list of edges: each line should contain two nonnegative integers, separated by whitespace: indicating an edge between the two nodes indicated by the given indices. Any lines starting with '#' are ignored. If there is anything after these first two integers, it is ignored as well.
The indices are assumed to be consecutive, i.e. starting at zero, with no nonnegative integers unused. If your indices start at 100000, the parsed graph will also have (orphaned) nodes for all integers from 0 to 100000.
You can find some examples packaged with the Nodes library. Files from the KONECT repository should also work out of the box.
The GML format is also supported with the switch --filetype gml
. This is not well tested, so your mileage may vary.
If your graph is large (in the order of millions of edges), you may need to use a disk-backed store. For this, the graph should be pre-loaded. This is done with the command
java -jar motive.jar --type preload --file data.txt
The graph will be loaded into a disk-backed database called graph.db
. You can then run an experiment with the type fast.disk
.
java -jar motive.jar --type fast.disk --file data.db
Make sure to use the -Xmx
argument (before the -jar
argument) to set the heap size as large as you can on both commands. For a node with 64 Gb of memory, we found that -Xmx56g
was the largest to give stable results.
You may still get out of memory errors during the computation of the code lengths (after the sampling phase) if the graph is large. This happens if a large number of links need to be rewritten to compute the motif code (in which case the motif is likely poor anyway). To combat this add the switch
--fast.max-rw 500000
to the command. If a motif requires more links to be rewritten than this amount, it will be skipped (although it may still succeed with a lower nunber of instances). A warning will be written to the error log every time this happens.
The command line tool produces its primary output as a collection of text files. For each (potential) motif found, one .edgelist file is produced. This file captures only the structure of the motif, not its labels in the original graph. That is, if the motif has the shape o -> o <- o
, then the corresponding edgelist file may look like
0 1
1 2
For each motif, there will be a CSV file containing its occurrences. If the motif has k nodes (i.e. k=3 in the example above), then each line in the occurrences file identifies a place where the motif occurs by providing a graph node integer (corresponding to the integers in the input file) for each node in the motif.
The files numbers.csv
contains the compression ratios for each motifs. The columns are, in order, the motif frequencies, the compression ratios for the ER model, compression ratios for the EL model, and (if the full
experiment is run) compression ratios for the DS model. In practice the EL model provides a good trade-off between speed and motif quality.
metadata.json
contains some additional information, mostly required for producing the correct plot.
The plots in the paper were produced using python scripts. Motive will copy these into the output directory and attempt to run them. This will fail if the correct dependencies are not installed. Here's a short recipe for getting the scripts to run:
- Install Anaconda (the python 3 version)
- run the command ```conda install matplotlib'''
- Go to the output directory and run ```python plot.synthetic.py''' (or whichever python file was copied to the output directory).
If this doesn't work, make a ticket on github describing what went wrong, and we'll try to help you out.
- The Nauty implementation, used for graph canonization is a re-implementation. It is not complete, and the real Nauty is likely much faster. If many samples of large motifs (eg. 12 nodes) are necessary, this will become the bottleneck. Up to 10 nodes, however, you should be able to take around a 25k 10-node samples per minute, even on commodity hardware.
- If you get a memory error, increase your heap space by adding
-Xmx3g
before-jar
. This sets the heap space to 3 gigabytes. Don't set the heap space to more memory than your machine has available (and leave at least 500m to 1000m for the system). If you still get memory errors, reduce the maximum motif size, the number of samples, or see if you can work with a smaller dataset (or move to a machine with more memory, of course).