[Model] Add BGNN example #2740
[Model] Add BGNN example #2740
Conversation
|
Thanks a lot! It's a brand new model contribution so I put an item in the model index README. Could you also add a README with instructions on how to download the dataset (once you have uploaded them)? |
|
@BarclayII @jermainewang I updated the code in the example and provided instructions how you can download datasets. Please let me know if something else is needed from me? |
examples/pytorch/bgnn/BGNN.py
Outdated
| ''' | ||
|
|
||
| :param graph: dgl graph | ||
| :param X: numpy matrix of input features. Can be composed of categorical features with missing values. |
There was a problem hiding this comment.
describe how missing values are represented in the matrix
examples/pytorch/bgnn/BGNN.py
Outdated
| :param logging_epochs: log every n epoch | ||
| :param metric_name: metric to use for early stopping | ||
| :param normalize_features: normalize original input features X (column wise) | ||
| :param replace_na: replace NA in X |
There was a problem hiding this comment.
I wonder if these options (normalize_features and replace_na) should be here. They look like data preprocessing steps that may go to the example code.
There was a problem hiding this comment.
Moved them to run.py for preprocessing.
|
Hey @jermainewang I updated the code documentation and also moved the preprocessing to run.py (that requires adding one additional parameter for fit function, as GBDT still want to operate on non-preprocessed features). That worked on my side. I hope it works for you too. |
|
@BarclayII do you have other comments? |
|
I'm good. Thanks! |
Description
A proposal of a new GNN model coming from ICLR 2021: https://openreview.net/forum?id=ebS5NUfoMKL
The model is implemented with DGL: https://github.com/nd7141/bgnn
Reviewers: @jermainewang @BarclayII
Discussion thread: #2603
This is initial proposal of the model. There are two files:
BGNN.pywhich implements BGNN model andrun.pywhich shows how to use it.I also include datasets I used in the paper for the sake of reproducibility (although I think that datasets should be included in DGL as a separate request).
Current results match those of the papers (although I have not done thorough hyperparameter tuning).
Here are images showing performance of BGNN, that is similar to the paper (https://openreview.net/forum?id=ebS5NUfoMKL):
I think
BGNN.pyshould be part of the DGL, in the sense that an end user can just doimport dgl.BGNNPredictor. On the other hand,run.pycan be kept in the example.A few notes:
BGNN.pyincludes as well plotting, which may not be needed for the final version.I am happy to answer your questions and discuss how to integrate BGNN in DGL in the optimal way. Please let me know your thoughts.