A Benchmark Dataset for Graph Regression

The repository is now organized around 3 workflows:

1. Learn and run examples: src/Tutorial - Start Here!/
2. Replicate paper results: src/replicate paper results/
3. Build a custom dataset: src/build your own regression dataset/

Repository map

graph_regression_datasets/
├── src/
│   ├── Tutorial - Start Here!/
│   │   ├── Tutorial.ipynb
│   │   ├── relsc_h.py
│   │   ├── relsc_m.py
│   │   ├── h_models.py
│   │   ├── hg_models.py
│   │   └── early_stopping.py
│   ├── replicate paper results/
│   │   ├── relsc_m_experiments.py
│   │   ├── relsc_m.py
│   │   └── dataset_statistics/
│   └── build your own regression dataset/
│       ├── FA-AST_java.py
│       └── mini_java/
│           ├── MiniTest.java
│           └── fa_ast_output.json
├── requirements.txt
└── Readme.md

Installation

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

Notes:

• torch-geometric must match your installed PyTorch version.
• If you use dataset-building scripts, install extra dependencies:
  • pip install javalang anytree

Dataset download

Dataset files are hosted on Zenodo:

• https://zenodo.org/records/18598713

Expected dataset folders:

graph_regression_datasets/
├── data/
│   ├── rdf.json
│   ├── dubbo.json
│   ├── H2.json
│   ├── hadoop.json
│   ├── systemds.json
│   └── ossbuilds.json
└── y_labels/
    ├── y_rdf.csv
    ├── y_dubbo.csv
    ├── y_H2.csv
    ├── y_hadoop.csv
    ├── y_systemds.csv
    └── y_ossbuilds.csv

Workflow 1: Start here (tutorial)

Main entry point:

• src/Tutorial - Start Here!/Tutorial.ipynb

Launch from repo root:

PYTHONPATH="src/Tutorial - Start Here!" jupyter lab "src/Tutorial - Start Here!/Tutorial.ipynb"

If your dataset folders are in repo root (data/, y_labels/), use these paths in loaders:

• x_folder="data"
• y_folder="y_labels"

Example:

from relsc_m import RelSCM

dataset = RelSCM(root="./data", project_name="rdf") # or dataset = RelSC(project_name="rdf")

split_idx = relsc_m_dataset1.get_idx_split()
train_loader = DataLoader(relsc_m_dataset1[split_idx['train']], batch_size=batch_size, shuffle=True)
val_loader   = DataLoader(relsc_m_dataset1[split_idx['val']],   batch_size=batch_size, shuffle=False)
test_loader  = DataLoader(relsc_m_dataset1[split_idx['test']],  batch_size=batch_size, shuffle=False)

Workflow 2: Replicate paper results (Type2)

Script:

• src/replicate paper results/relsc_m_experiments.py

Run from repo root, ensure to have all the libraries that are necessary:

PYTHONPATH="src/Tutorial - Start Here!" python3 "src/replicate paper results/relsc_m_experiments.py"

What it does:

• runs all configured RelSCM datasets,
• runs two RelSCM models (HeteroGraphConv, HeteroTransformer),
• evaluates multiple seeds,
• exports CSV tables (for example Table3_TEST_MAE.csv).

Static plots and stats used in the paper are in:

• src/replicate paper results/dataset_statistics/

Workflow 3: Build your own regression dataset

Script:

• src/build your own regression dataset/FA-AST_java.py

FA-AST_java.py parses Java files and builds FA-AST graph payloads.

Run on your own Java folder:

python3.11 "src/build your own regression dataset/FA-AST_java.py" \
  --input-dir "/path/to/java/files" \
  --output-json "fa_ast_output.json"

Included mini example (already in repo):

• Input file: src/build your own regression dataset/mini_java/MiniTest.java
• Generated graph: src/build your own regression dataset/mini_java/fa_ast_output.json

Regenerate the included sample graph:

python3.11 "src/build your own regression dataset/FA-AST_java.py" \
  --input-dir "src/build your own regression dataset/mini_java" \
  --output-json "src/build your own regression dataset/mini_java/fa_ast_output.json"

Quick verification test:

python3.11 - <<'PY'
import json

path = "src/build your own regression dataset/mini_java/fa_ast_output.json"
with open(path, "r", encoding="utf-8") as f:
    data = json.load(f)

print("graphs:", len(data))
assert len(data) > 0, "No graphs generated"

sample_key, sample_value = next(iter(data.items()))
payload, ast_len = sample_value
x, edge_index, edge_attr = payload
print("sample:", sample_key)
print("ast_len:", ast_len, "nodes:", len(x), "edges:", len(edge_index[0]), "edge_attr:", len(edge_attr))
assert len(x) == ast_len, "Node count mismatch"
assert len(edge_index[0]) == len(edge_attr), "Edge/attribute mismatch"
print("FA-AST smoke test: OK")
PY

Common path note

Folder names under src/ contain spaces. Quote paths in shell commands:

python3 "src/replicate paper results/relsc_m_experiments.py"

Citation

Samoaa, P., Vukojevic, M., Haghir Chehreghani, M., & Longa, A. (2026). Broadening the Scope of Graph Regression: Introducing a Dataset with Multiple Representation Settings [Data set]. Zenodo. https://doi.org/10.5281/zenodo.18598713