inf554_challenge

In order to run all the pre-processing algorithms and the models, please download abstracts.txt, author_papers.txt, train.csv, test.csv and coauthorship.edgelist in the data folder. These files are available on the following link: https://www.kaggle.com/c/inf554-2021/data. Then, create a virtual environment, activate it and install the required libraries with

python3 -m venv .venv
source .venv/bin/activate
pip3 install -r requirements.txt

Once it is done, you can start pre-processing the data to create the necessary embeddings.

Pre-processing

The first step of pre-processing is to extract the filtered vocabulary from the abstracts.txt file. This can be done by running

cd code/embeddings
python3 vocabulary-processing.py

Then, you can extract full sentences (for supervised embeddings including SBERT and InferSent) from the abstract.txt file with

cd code/embeddings
python3 sentence-builder.py

To create mean word2vec-like embeddings (word2vec, GloVe, fastText), you can run the provided .sh script.

cd code/embeddings
./run-meanvec-parser.sh

To create InferSent embeddings (DEPRECATED)

cd code/embeddings
./prepare-infersent-parser.sh
./run-infersent-parser.sh

To create SBERT embeddings

cd code/embeddings
./run-transformer-parser.sh

Finally, to create node2vec embeddings

cd code/embeddings
./create-node2vec.sh

Finally, to merge embeddings and create the full features that will be stored in data/author-embeddings

cd code/embeddings
./merge-embeddings.sh

Models

Multilayer perceptron (MLP)

To perform k-fold cross validation with different sets of hyperparameters on a given author-embeddings, do

cd code/MLP
python3 MLP-k-fold.py <author-embeddings>

The logs obtained with SPECTER transformer-based embeddings reduced to 128 dimensions and concatenated with deepwalk embeddings of size 128 are provided in data/logs/allenai-specter-pca-128-deepwalk-128-logs.txt.

To perform training with a multilayer perceptron with chosen parameters

cd code/MLP
python3 MLP-pipeline.py <author-embeddings> <n_hidden1> <n_hidden2> <n_epochs> <dropout_p> <learning-rate> <test-output> [<plot-data>]

If you exported plot data, you can display it thanks to MLP-figure.py

cd code/MLP
python3 MLP-figure.py <output-figure> <input-losses> [<other-input-losses> ...]

XG-Boost

To perform k-fold cross validation with different sets of hyperparameters on a given author-embeddings, do

cd code/xgboost
python3 xgboost-cross_validation.py <author-embeddings> <n_estimators> <max_depth> <learning-rate> <subsample> <k-fold>

To perform training with chosen parameters

cd code/xgboost
python3 xgboost-submission.py <author-embeddings> <n_estimators> <max_depth> <learning-rate> <subsample> <k-fold> <submission_csv>

Graph Neural Networks

Once you have an abstract embeddings file ready, execute the following instructions to compute node embeddings (not taking graph node metrics into consideration yet):

cd code/GNN
python3 abstracts_to_author.py <abstract-embeddings-file> <abstract-embeddings-size>

To add node metrics to the author embeddings file

cd code/GNN
python3 graph_metrics_features.py <author-embeddings-file> <abstract-embeddings-size>

Choose your model among GCN, GAT and GraphSAGE and execute one of those three lines:

cd code/GNN
python3 pipeline_GCN.py <author-full-embeddings-file> <author-full-embeddings-size> <size-layer1> <size-layers2-3> <dropout-rate>
python3 pipeline_GAT.py <author-full-embeddings-file> <author-full-embeddings-size> <size-layer1> <size-layers2-3> <dropout-rate>
python3 pipeline_GraphSAGE.py <author-full-embeddings-file> <author-full-embeddings-size> <size-layer1> <size-layers2-3> <dropout-rate> <aggregation-function>

Post-processing

Execute the following instruction for post-processing

cd code/GNN
python3 postprocessing.py <predictions-timestamp>