Skip to content

Latest commit

 

History

History
97 lines (85 loc) · 4.94 KB

README.md

File metadata and controls

97 lines (85 loc) · 4.94 KB

Supervised Graph Neural Network (aka supGNN) paired with downstream XGBoost model

Introduction

Recommendation datasets are often modelled as graph datasets because of high predictive prowess of Graph Neural Networks. Here, we create a graph dataset by using either a column or a group of columns from ShareChat dataset and train a supervised GraphSage model on it with an emphasis on ensuring no information from test data is used during training and all test examples are inferred independently i.e. no two test edges can see eachother.


Self-supervised bipartite GNN pipeline.

!Important

This is an end-to-end README, Train and Inference instructions are both included.

Set up

  1. Create working directory and subfolders for storing intermediate data files
mkdir data && cd data
mkdir supGNN_graph_data sharechat_recsys2023_data
cd supGNN_graph_data 
mkdir full_graph train_graph test_graph
  1. Download ShareChat dataset inside sharechat_recsys2023_data folder. It should contain train and test folders with the given csv files.

Running pipeline

Section 1 : employ Supervised GNN to generate node embeddings

Please see outline for section 1 in following diagram - Supervised GNN to generate node embeddings

Convert tabular data to edge list

We model Group dataset into nodes based on features f_6 for role-1 and f_2, f_4, and f_16 for role-2.

  1. Convert tabular training data into training graph Using only ./sharechat_recsys2023_data/train folder, we create ./data/supGNN_graph_data/train_graph/train_edges.csv.gz which will be used for supGNN train/localdisk/akakne/recsys2023/assets/recsys_supgnn.jpging later.
python3 0_train/2_supGNN/convert_train_tabular_data_to_edge_list.py
  1. Create full graph for supGNN inference Using both ./sharechat_recsys2023_data/train and ./sharechat_recsys2023_data/test folders, we create ./supGNN_graph_data/full_edges.csv.gz which will be used for supGNN inference later.
python3 1_inference/2_supGNN/convert_full_tabular_data_to_edge_list.py

Create CSVDataset from edge list

  1. Convert training graph into train CSVDataset (for GNN training)
python3 0_train/2_supGNN/convert_train_edge_list_to_CSVDataset.py

Please see newly created files inside ./data/supGNN_graph_data/train_graph/recsys_graph folder.

  1. Convert full graph into full CSVDataset (for GNN inference)
python3 1_inference/2_supGNN/convert_full_edge_list_to_CSVDataset.py

Please see newly created files inside ./data/supGNN_graph_data/full_graph/recsys_graph folder.

Run supervised Graph Neural Network to generate node embeddings

  1. Train supervised GNN on train graph using ./data/supGNN_graph_data/train_graph/recsys_graph.
python3 0_train/2_supGNN/train_supervised_graphsage.py

Above script saves best model (across epochs) as ./data/supGNN_graph_data/train_graph/model.pt
IMPORTANT NOTE : we are using only train graph in this step i.e. GNN model will never see any test data during training.

  1. Run inference on full graph using saved GNN model 
python3 1_inference/2_supGNN/infer_supervised_graphsage.py

IMPORTANT NOTE : to ensure indepedent testing for all test edges, we disable message passing through test edges i.e. any test edge e will see only train and validation edges around it's source and destination nodes. Edge e will never see another test edge e'. This is done by setting the probablity of picking test edges as neighbours to 0. For more details see our inference code in file ./1_inference/2_supGNN/infer_supervised_graphsage.py line 112 and 20 respectively. We also handle unseen nodes in this script.

Section 2 : Create GNN-boosted features and run XGBoost for binary classification

Please see outline for section 2 in following diagram - Create GNN-boosted features and run XGBoost for binary classification

Map GNN embeddings to original features

python3 1_inference/2_supGNN/map_node_emb_to_edge_list.py

this script will map GNN-generated node embeddings to their respective edges and save GNN-boosted features for train and test split separately.

Merge supGNN data with Feature Engineered data

  1. Merge train supGNN data with train FE data
python3 0_train/2_supGNN/merge_train_FE_and_train_supgnn_data.py
  1. Merge test supGNN data with test FE data
python3 1_inference/2_supGNN/merge_test_FE_and_test_supgnn_data.py

Run XGBoost on merged data

  1. Train XGB on merged train data
python3 0_train/2_supGNN/train_xgb.py
  1. Infer XGB on merged test data
python3 1_inference/2_supGNN/infer_xgb.py

Finally, we get ./data/supGNN_graph_data/test_graph/submission.csv as our final submission file which we upload to the leaderboard for evaluation on test set.