Skip to content
Branch: master
Find file History
BarclayII and jermainewang [Sampler] Metapath sampler for metapath2vec (#861)
* metapath sampler

* lint & fixes

* lint x2

* lint x3

* fix windows

* remove max_cycle argument

* add todo note
Latest commit e3921d5 Sep 27, 2019

README.md

Metapath2vec

Dependencies

  • PyTorch 1.0.1+

How to run the code

Run with the following procedures:

1, Run sampler.py on your graph dataset. Note that: the input text file should be list of mappings so you probably need to preprocess your graph dataset. Files with sample format are available in "net_dbis" file. Of course you could also use your own metapath sampler implementation.

2, Run the following command:

python metapath2vec.py --download "where/you/want/to/download" --output_file "your_output_file_path"

Tips: Change num_workers based on your GPU instances; Running 3 or 4 epochs is actually enough.

Tricks included in the implementation:

1, Sub-sampling;

2, Negative Sampling without repeatedly calling numpy random choices;

Performance and Explanations:

Venue Classification Results for Metapath2vec:

Metric 5% 10% 20% 30% 40% 50% 60% 70% 80% 90%
Macro-F1 0.3033 0.5247 0.8033 0.8971 0.9406 0.9532 0.9529 0.9701 0.9683 0.9670
Micro-F1 0.4173 0.5975 0.8327 0.9011 0.9400 0.9522 0.9537 0.9725 0.9815 0.9857

Author Classfication Results for Metapath2vec:

Metric 5% 10% 20% 30% 40% 50% 60% 70% 80% 90%
Macro-F1 0.9216 0.9262 0.9292 0.9303 0.9309 0.9314 0.9315 0.9316 0.9319 0.9320
Micro-F1 0.9279 0.9319 0.9346 0.9356 0.9361 0.9365 0.9365 0.9365 0.9367 0.9369

Note that:

Testing files are available in "label 2" file;

The above are results listed in the paper, in real experiments, exact numbers might be slightly different:

1, For venue node classification results, when the size of the training dataset is small (e.g. 5%), the variance of the performance is large since the number of available labeled venues is small.

2, For author node classification results, the performance is stable since the number of available labeled authors is huge, so even 5% training data would be sufficient.

3, In the test.py, you could change experiment times you want, especially it is very slow to test author classification so you could only do 1 or 2 times.

You can’t perform that action at this time.