8th place solution.
Competition page:
https://www.automl.ai/competitions/3
https://www.4paradigm.com/competition/kddcup2020
Class Model from model.py implements the API required by the evaluation system.
Example of local running:
docker run --gpus=0 --shm-size=30G -it --rm -v "$(pwd):/app/autograph" -v /tmp/pipdocker:/root/.cache/pip -w /app/autograph nehzux/kddcup2020:v2python starting_kit/run_local_test.py --dataset_dir=./starting_kit/data/demo/ --code_dir=./src/
Please refer to the official documentation for the detailed interface description.
Let's just train several different architectures for the node classification task:
- TagConv - Topology Adaptive Graph Convolutional Networks
- SageConv - Inductive Representation Learning on Large Graphs
- GraphConv - Semi-Supervised Classification with Graph Convolutional Networks
- SGConv - Simplifying Graph Convolutional Networks
And then average the results of the top performing models (evaluated on the validation).