This project implements the DynamicTriad algorithm proposed in [1], which is a node embedding algorithm for undirected dynamic graphs.
This project is implemented primarily in Python 2.7, with some c/c++ extensions written for time efficiency.
Though the program falls back to pure Python implementation if c/c++ extensions fail to build, we DISCOURAGE you from using these code because they might have not been actively maintained and properly tested.
The c/c++ code is ONLY compiled and tested with standard GNU gcc/g++ compilers (with c++11 and OpenMP support), and other compilers are explicitly disabled in our build scripts. If you have to use another compiler, modifications on build scripts are required.
- Boost.Python. Version 1.54.0 has been tested. You can find instructions to install from source here.
- CMake. Version >= 2.8 required. You can find installation instructions here.
- Eigen 3. Version 3.2.8 has been tested, and later versions are expected to be compatible. You can find installation instructions here.
- Python 2.7. Version 2.7.13 has been tested. Note that Python development headers are required to build the c/c++ extensions.
- graph-tool. Version 2.18 has been tested. You can find installation instructions here.
- TensorFlow. Version 1.1.0 has been tested. You can find installation instructions here. Note that the GPU support is ENCOURAGED as it greatly boosts training efficiency.
- Other Python modules. Some other Python module dependencies are listed in
requirements.txt, which can be easily installed with pip:pip install -r requirements.txt
Although not necessarily mentioned in all the installation instruction links above, you can find most of the libraries in the package repository of a regular Linux distribution.
Before building the project, we recommend switching the working directory to the project root directory. Assume the project root is at <dynamic_triad_root>, then run command
cd <dynamic_triad_root>
Note that we assume <dynamic_triad_root> as your working directory in all the commands presented in the rest of this documentation.
A building script build.sh is available in the root directory of this project, simplifying the building process to executing a single command
bash build.sh
Before running the actual building commands, the script requires you to configure some of the environment variables. You can either use the default values or specify your custom installation paths for certain libraries. For example,
PYTHON_LIBRARY? (default: /usr/lib64/libpython2.7.so.1.0, use a space ' ' to leave it empty)
PYTHON_INCLUDE_DIR? (default: /usr/include/python2.7, use a space ' ' to leave it empty)
EIGEN3_INCLUDE_DIR? (default: /usr/include, use a space ' ' to leave it empty)
BOOST_ROOT? (default: , use a space ' ' to leave it empty) ~/boost_1_54_1
If everything goes well, the build.sh script will automate the building process and create all necessary binaries.
Note that the project also contains some Cython modules, however, they will be automatically built as soon as the module is imported if the environment is ready.
A test script scripts/test.py is available, run
python scripts/test.py
to see if everything is fine with building.
Given a sequence of undirected graphs, each for a time step, this program can be used to compute a real-valued vector for each vertex at each time step.
The input is expected to be a directory containing N input files named 0, 1, 2..., where N is the length of the graph sequence. Each file contains an adjacency list of the corresponding graph, and the adjacency list consists of multiple lines, each in the format:
<from_node_name> [<to_node_name1> <weight1> [<to_node_name2> <weight2> ...] ]
where x_node_name can be any ascii string without white space characters in it, and weight are float or integer values. The line describes edges from from_node_name to to_node_name1 and to_node_name2 with weight weight1 and weight2 respectively.
Note that:
- The graph is expected to be undirected, however, it should be presented in a directed format. That is, if there is an edge (u, v, w), its reciprocal edge (v, u, w) must also exists in the adjacency list.
- The vertex set should be same for all graphs, if a vertex is missing in a certain graph, simply present it as an isolated vertex.
- If a vertex has no outbound vertices, you should write a line with only the
from_node_name, instead of ignoring this vertex. - If the graph is unweighted, place a
1.0for allweightplaceholders, rather than ignoring all weights in the adjacency list. - Loopback edges (u, u, w) will be ignored when the adjacency list is loaded.
The program outputs to a directory creating N files named 0.out, 1.out, 2.out, ..., each corresponds to an input file (time step). Each output file contains V lines, where V is the number of vertices in each graph. And each line is in format:
<node_name> <r1> <r2> ... <rK>
where <node_name> is the name of the vertex defined in the input files, which is followed by K real values, i.e. the K-length embedding vector for vertex <node_name> at the corresponding time step.
Now that the input data is ready, the main script will be called to compute dynamic vertex embeddings. Following the assumption that the current working directory is <dynamic_triad_root>, the help information of the main script can be obtain by executing command
python . -h
usage: . [-h] [-I NITERS] -d DATAFILE [-b BATCHSIZE] -n NSTEPS
[-K EMBDIM] [-l STEPSIZE] [-s STEPSTRIDE] -o OUTDIR
[--cachefn CACHEFN] [--lr LR] [--beta BETA [BETA ...]]
[--negdup NEGDUP] [--validation VALIDATION]
optional arguments:
-h, --help show this help message and exit
-I NITERS, --niters NITERS
number of optimization iterations (default: 10)
-d DATAFILE, --datafile DATAFILE
input directory name (default: None)
-b BATCHSIZE, --batchsize BATCHSIZE
batchsize for training (default: 5000)
-n NSTEPS, --nsteps NSTEPS
number of time steps (default: None)
-K EMBDIM, --embdim EMBDIM
number of embedding dimensions (default: 48)
-l STEPSIZE, --stepsize STEPSIZE
size of of a time steps (default: 1)
-s STEPSTRIDE, --stepstride STEPSTRIDE
interval between two time steps (default: 1)
-o OUTDIR, --outdir OUTDIR
output directory name (default: None)
--cachefn CACHEFN prefix for data cache files (default: None)
--lr LR initial learning rate (default: 0.1)
--beta-smooth BETA_SMOOTH
coefficients for smooth component (default: 0.1)
--beta-triad BETA_TRIAD
coefficients for triad component (default: 0.1)
--negdup NEGDUP neg/pos ratio during sampling (default: 1)
--validation VALIDATION
link_prediction, link_reconstruction, node_classify,
node_predict, none (default: link_reconstruction)
Some of the arguments may require extra explanation:
--beta-smooth/--beta-triad, two hyper parameters used in the model, see reference [1] for details about the hyper parameters of DynamicTriad. Empirically, the hyper parameters need to be tuned in order to achieve the best performance, and the best choice depends on the task and the stability of the target dynamic network.-l/--stepsizeand-s/--stepstride, see Time Model for details.--cachefn, sometimes you find that the data preprocessing becomes intolerably time consuming (see Time Model), and a solution is to specify--cachefnso that the program creates or uses a cache file of the preprocessed data. The cache file consists of two parts -- a file named<--cachefile>.cacheas well as a file named<--cachefile>.cache.args. If you have changed your configuration for preprocessing, remove<--cachefile>.cache.argsand the cache will be regenerated.--validation, the four tasks available for validation are as defined in [1], please refer to the paper for details.
We include a toy data set in the data directory, namely data/academic_toy.pickle, which is a subset of Academic data set in [1] stored using Python pickle module. See Data Sets for more details.
A demo script is available as scripts/demo.sh, which primarily does three things:
- Call
scripts/academic2adjlist.pyto convert the toy data to the input format described in Input Format. - Call the main script to compute the vertex embeddings and save them to
outputdirectory. - Call
scripts/stdtest.pyto experiment on standard tasks described in paper [1].
In the demo script, you can find an example for the standard usage of the main script, as well as hints for the usage of the other two scripts, if you are interested in them.
To run the demo, execute command
bash scripts/demo.sh
TL;DR: If you would like the main script to treat your graphs exactly as they are specified in your input files, please leave the arguments -l and -s to their default values.
For flexibility, a part of the data preprocessing functionalities are included into our main script. Specifically, if we call each graph file in the input directory a unit graph, our main script provides interfaces to create the graph for each time step out of these unit graphs.
Before describing this preprocessing step, we shall first define a time step. According to our assumption, a time step consists of <stepsize> consecutive unit graphs, where <stepsize> is a constant value shared across all time steps. There are <stepstride> - 1 unit graphs between the leading unit graphs of two adjacent time steps, where <stepstride> is also a constant value. For example, we set <stepsize>=4 and <stepstride>=2 in our demo script, as a result, the time steps are:
time step #1: unit graph 0 -- unit graph 3
time step #2: unit graph 2 -- unit graph 5
time step #3: unit graph 4 -- unit graph 7
...
Once <stepsize> and <stepstride> are given, each time step now corresponds to a subsequence of unit graphs, and the graph for this time step is created by merging these unit graphs, i.e. by summing up weights for the same edge.
Note that if you set both <stepsize> and <stepstride> to 1, the graphs will be used as is specified in the input directory. If the merging operation is found very time expensive, specifying a <--cachefile> avoids re-merging everytime you run the script, as long as the data configuration is kept unchanged.
One out of the three data sets reported in [1] -- the Academic Data Set -- was made public by AMiner, which consists of information about papers published in a recent few decades. We keep only those papers published between 1980 and 2015 (included), and we remove from the data those researchers with less than 15 publications in total and conferences with less than 20 participants in total, so that the resulting dynamic network becomes more stable.
In this data set, labels are extracted for each researcher indicating the research fields he/she focuses on. We manually specify a set of representing conferences for each research field, and try to find out for a researcher in which field he/she publishes most of his/her work, given a certain time step.
A toy data is included in this project as data/academic_toy.pickle, which was originally the ACM-Citation-network V8 data set from AMiner, and was preprocessed as we describe above, with the only difference that the vertices are further sampled to a limited size of 2000. And our full preprocessing result can be downloaded here.
Update: For those who are interested in the academic dataset and wish to avoid the bothering building process, the dataset in clean format is released here (Please cite the original publisher of the data if you wish use the dataset). See readme.txt in the package for detailed information, and feel free to contact me if there are anything wrong or unclarified in the data.
As reported in [1], the performance of DynamicTriad on Academic Data Set with embedding dimension set to 48 is:
| F1-score on Academic | Vertex Classification | Link Reconstruction | C.Link Reconstruction |
|---|---|---|---|
| DeepWalk | 0.630 | 0.694 | 0.702 |
| node2vec | 0.359 | 0.574 | 0.611 |
| Temporal Network Embedding | 0.625 | 0.974 | 0.899 |
| DynamicTriad | 0.704 | 0.985 | 0.925 |
| F1-score on Academic | Vertex Prediction | Link Prediction | C.Link Prediction |
|---|---|---|---|
| DeepWalk | 0.591 | 0.612 | 0.674 |
| node2vec | 0.355 | 0.548 | 0.617 |
| Temporal Network Embedding | 0.596 | 0.772 | 0.889 |
| DynamicTriad | 0.671 | 0.836 | 0.924 |
Please refer to [1] for more information about our experiments, where you can find the definition of tasks, the experimental settings, the description of unpublished data sets and the full results of our experiments.
[1] Zhou, L; Yang, Y; Ren, X; Wu, F and Zhuang, Y, 2018, Dynamic Network Embedding by Modelling Triadic Closure Process, In AAAI, 2018
@inproceedings{zhou2018dynamic,
title = "{Dynamic Network Embedding by Modelling Triadic Closure Process}",
author = {{Zhou}, L. and {Yang}, Y. and {Ren}, X. and {Wu}, F. and {Zhuang}, Y.},
booktitle={AAAI},
year = 2018,
}