TabGraphs: A Benchmark and Strong Baselines for Learning on Graphs with Tabular Node Features

About the datasets

The proposed TabGraphs benchmark can be downloaded via our Zenodo record. It is necessary to put the compressed .zip files into the datasets directory. To unzip a dataset <dataset_name>, one can run unzip <dataset_name> in their terminal.

In each dataset subfolder, we provide the following files:

• features.csv — node features
• targets.csv — node targets
• edgelist.csv — list of edges in graph
• train_mask.csv, valid_mask.csv, test_mask.csv — split masks

Besides that, we put info.yaml with the necessary information about dataset:

• dataset_name — dataset name
• task — prediction task
• metric — metric used to evaluate predictive performance
• num_classes — number of classes, if applicable
• has_unlabeled_nodes — whether dataset has unlabaled nodes
• has_nans_in_num_features — whether dataset has NaNs in numerical features
• graph_is_directed — whether graph is directed
• graph_is_weighted — whether graph is weighted (if true, then edgelist.csv has 3 columns instead of 2)
• target_name — target name
• num_feature_names — list of numerical feature names
• cat_feature_names — list of categorical feature names
• bin_feature_names — list of binary feature names

Note! The proposed TabGraphs benchmark is released under the CC BY 4.0 International license.

About the source code

In source directory, one can also find the source code for reproducing experiments in our paper. Note that only gnns subfolder contains our original code, while subfolders bgnn, ebbs and tabular are taken from open sources and adapted to make them consistent with our experimental setup.

Further, we provide the original sources:

• tabular — github.com/yandex-research/tabular-dl-tabr
• bgnn — github.com/nd7141/bgnn
• ebbs — github.com/JiuhaiChen/EBBS

The only changes that were made in the original repositories are related to the logging of experimental results and the metrics used for validation.

How to reproduce results

1. Run notebook notebooks/prepare-graph-augmentation.ipynb to prepare graph-based feature augmentations (NFA) that can be used by tabular baselines from tabular.
2. Run notebook notebooks/prepare-node-embeddings.ipynb to prepare optional DeepWalk embeddings (DWE) for the proposed datasets that can further improve predictive performance.
3. Run notebook notebooks/convert-graph-datasets.ipynb to convert the provided graph datasets (probably with NFA and/or DWE) into the format required by tabular baselines and specialized models bgnn and ebbs.
4. Run experiments according to the instructions provided in the corresponding directories.

Note! The source code for tabular baselines and bgnn model is distributed under the MIT license, and our code for gnns is also released under the same MIT license.