r_unimp

R-UNIMP for MAG240M using PGL

The code is about 《R-UNIMP: SOLUTION FOR KDDCUP 2021 MAG240M-LSC》.

@article{shirunimp,
  title={RUnimp: SOLUTION FOR KDDCUP 2021 MAG240M-LSC},
  author={Shi, Yunsheng and Team, PGL and Huang, Zhengjie and Li, Weibin and Su, Weiyue and Feng, Shikun}
}

Installation Requirements

ogb==1.3.0
torch==1.7.0
paddle==2.0.0
pgl==2.1.2

Running Preprocessing Script

Get the features:

python dataset/sage_author_x.py
python dataset/sage_institution_x.py
python dataset/sage_author_year.py
python dataset/sage_institution_year.py
python dataset/sage_all_data.py
python dataset/merge_m2v_embed.py

This will give you the following files:

• author.npy: The author features, preprocessed by averaging the neighboring paper features.
• institution_feat.npy: The institution features, preprocessed by averaging the neighboring author features.
• author_year.npy: The author year, preprocessed by averaging the neighboring paper years.
• institution_year.npy The institution years, preprocessed by averaging the neighboring author years.
• full_feat.npy: The concatenated author, institution, and paper features.
• all_feat_year.npy: The concatenated author, institution, and paper years.
• m2v_embed.npy: The m2v embed. you can get it from here (password: 0mr0), using dataset/merge_m2v_embed.py to merget them together.
• paper_to_paper_symmetric_pgl_split: The paper_to_paper PGL graph.
• paper_to_author_symmetric_pgl_split_src: The author_to_paper PGL graph.
• paper_to_author_symmetric_pgl_split_dst: The paper_to_author PGL graph.
• institution_edge_symmetric_pgl_split_src: The author_to_institution PGL graph.
• institution_edge_symmetric_pgl_split_dst: The institution_to_author PGL graph.

Get the Metapath2vec embedding:

We get metapath2vec embeddings following https://github.com/PaddlePaddle/PGL/tree/static_stable/examples/metapath2vec

Get the new validation split:

python split_valid.py

Then, you will save the new cross validation data in follow dir:

• ./valid_64

Running Multi-GPU Training Script

run_r_unimp_train.sh

Running Multi-GPU Inferring Script

run_r_unimp_infer.sh

This will give you R_UNIMP value in the performance table below

Running Post Process Script

1. Construct the coauthor graph

# Constructed Co-author Graph

python construct_coauthor_graph.py

2. Arange all the validation and test prediction file as following

./result/model1
             \_   all_eval_result.npy  # concatenate all validation output
             \_   test_0.npy           # Prediciton for Fold-0 model 
             \_   test_1.npy           # Prediciton for Fold-1 model 
             \_   test_2.npy           # Prediciton for Fold-2 model 
             \_   test_3.npy           # Prediciton for Fold-3 model 
             \_   test_4.npy           # Prediciton for Fold-4 model 
             \_   valid_0.npy          # validation-id for Fold-0
             \_   valid_1.npy          # validation-id for Fold-1
             \_   valid_2.npy          # validation-id for Fold-2
             \_   valid_3.npy          # validation-id for Fold-3
             \_   valid_4.npy          # validation-id for Fold-4

./result/model2
             \_   all_eval_result.npy  # concatenate all validation output
             \_   test_0.npy           # Prediciton for Fold-0 model 
             \_   test_1.npy           # Prediciton for Fold-1 model 
             \_   test_2.npy           # Prediciton for Fold-2 model 
             \_   test_3.npy           # Prediciton for Fold-3 model 
             \_   test_4.npy           # Prediciton for Fold-4 model 
             \_   valid_0.npy          # validation-id for Fold-0
             \_   valid_1.npy          # validation-id for Fold-1
             \_   valid_2.npy          # validation-id for Fold-2
             \_   valid_3.npy          # validation-id for Fold-3
             \_   valid_4.npy          # validation-id for Fold-4

3. Runing Post-Smoothing

model_name=model1

# set alpha = 0.8 and smoothing for each fold
python post_smoothing.py 0.8 0 ${model_name} 
python post_smoothing.py 0.8 1 ${model_name} 
python post_smoothing.py 0.8 2 ${model_name}
python post_smoothing.py 0.8 3 ${model_name} 
python post_smoothing.py 0.8 4 ${model_name} 


# merge result and generate ./result/${model_name}_diff0.8/all_eval_result.npy
python merge_result.py ${model_name}_diff0.8

4. Run ensemble

# This will automatically ensemble results from ./result/ and generate y_pred_mag240m.npz 
python ensemble.py

This will give you R_UNIMP_POST value in the performance table below

Performance

Model	Valid ACC
R_UNIMP	0.7715
R_UNIMP_POST	0.7729
Final Ensemble (30)	0.7773