IntelliGraphs is a collection of benchmark datasets specifically for use in benchmarking generative models for knowledge graphs. You can learn more about it in our preprint: IntelliGraphs: Datasets for Benchmarking Knowledge Graph Generation. The Python package provides easy access to the datasets, along with pre- and post-processing functions, baseline models, and evaluation tools for benchmarking new models.

IntelliGraphs can be installed using either pip or conda. Dependencies be automatically installed during the installation process.

pip install intelligraphs         # Standard pip
uv pip install intelligraphs     # Using UV (faster)

conda install -c thiv intelligraphs

After installation, you can verify that IntelliGraphs has been successfully installed by running the following command in your Python environment:

python -c "import intelligraphs; print(intelligraphs.__version__)"

It is recommended to use the latest version. If you don't have the latest version (check badge above), please ensure to update your installation before using it:

pip install --upgrade intelligraphs  # or conda install -c thiv intelligraphs --upgrade

The datasets required for this project can be obtained either manually or automatically through IntelliGraphs Python package.

The datasets are hosted on Zenodo: https://doi.org/10.5281/zenodo.14787483

You can download the datasets and extract the files to your preferred directory.

To download, verify, and extract datasets automatically, use:

python -m intelligraphs.data_loaders.download

This command will download all IntelliGraphs datasets, verify their integrity using MD5 checksums, and then extract them into the .data directory in your current working directory.

The DataLoader class is a utility for loading IntelliGraphs datasets, simplifying the process of accessing and organizing the data for machine learning tasks. It provides functionalities to download, extract, and load the datasets into PyTorch tensors.

1. Instantiate the DataLoader:

from intelligraphs import DataLoader
data_loader = DataLoader(dataset_name='syn-paths')

2. Load the Data:

train_loader, valid_loader, test_loader = data_loader.load_torch(
    batch_size=32,
    padding=True,
    shuffle_train=False,
    shuffle_valid=False,
    shuffle_test=False
)

3. Access the Data:

for batch in train_loader:
    # Perform training steps with the batch

for batch in valid_loader:
    # Perform validation steps with the batch

for batch in test_loader:
    # Perform testing steps with the batch

This generator creates path graphs where each node represents a city in the Netherlands and each edge represents a mode of transport (cycle_to, drive_to, train_to).

• Entities: Dutch cities
• Relations: Modes of transport between cities
• Use case: Structural learning

This generator creates graphs representing academic roles, timelines, and people. The nodes represent individuals, roles, and years, and the edges represent relationships like has_name, has_role, start_year, and end_year.

• Entities: Names, roles, years
• Relations: Relationships between academic roles and timeframes
• Use case: Basic temporal reasoning and type checking

This generator creates graphs where nodes represent countries, languages, and cities, and edges represent relationships like spoken_in, part_of, and same_as.

• Entities: Countries, languages, cities
• Relations: Geographical and linguistic relationships
• Use case: Type checking

Each generator class inherits from BaseSyntheticDatasetGenerator and can be customized by overriding methods or adjusting parameters. The base class provides utility methods for splitting datasets, checking for unique graphs, and visualizing graphs.

To create a new dataset generator, simply create a new class that inherits from BaseSyntheticDatasetGenerator and implement the sample_synthetic_data method to define your dataset's logic.

class MyCustomDatasetGenerator(BaseSyntheticDatasetGenerator):
    def sample_synthetic_data(self, num_graphs):
        # Implement your custom logic here
        pass

You can generate synthetic datasets by running the corresponding script for each generator. Each generator allows customization of dataset size, random seed, and other parameters.

python intelligraphs/generator/synthetic/synpaths_generator.py --train_size 60000 --val_size 20000 --test_size 20000 --num_edges 3 --random_seed 42 --dataset_name "syn-paths"
python intelligraphs/generator/synthetic/syntypes_generator.py  --train_size 60000 --val_size 20000 --test_size 20000 --num_edges 3 --random_seed 42 --dataset_name "syn-types"
python intelligraphs/generator/synthetic/syntipr_generator.py --train_size 50000 --val_size 10000 --test_size 10000 --num_edges 3 --random_seed 42 --dataset_name "syn-tipr"

Every dataset comes with a set of rules that describe the nature of the graphs. The ConstraintVerifier class includes a convenient method called print_rules() that allows you to view all the rules and their descriptions in a clean and organized format.

To use the print_rules() method, simply instantiate a subclass of ConstraintVerifier, such as SynPathsVerifier, and then call the print_rules() method on that instance to list the logical rules for a given dataset.

from intelligraphs.verifier.synthetic import SynPathsVerifier

# Initialize the verifier for the syn-paths dataset
verifier = SynPathsVerifier()

# Print the rules and their descriptions for the syn-paths dataset
verifier.print_rules()

When you call print_rules(), you'll get a formatted list of all the rules along with their corresponding descriptions. For example:

List of Rules and Descriptions:
        -> Rule 1:
           FOL: ∀x, y, z: connected(x, y) ∧ connected(y, z) ⇒ connected(x, z)
           Description: Ensures transitivity. If x is connected to y, and y is connected to z, then x should be connected to z.
        -> Rule 2:
           FOL: ∀x, y: edge(x, y) ⇒ connected(x, y)
           Description: If there's an edge between two nodes x and y, then x should be connected to y.
        ...

Our baseline models are also available through the Python API. You can find them inside baseline_models class.

To import the Uniform Baseline model:

from intelligraphs.baseline_models import UniformBaseline

To import the Knowledge Graph Embedding (KGE) models:

from intelligraphs.baseline_models.knowledge_graph_embedding_model import KGEModel

To recreate our experiments, we recommend using a fresh virtual environment with Python 3.10 installed.

pip install -e .  # or: pip install intelligraphs  # or: conda install -c thiv intelligraphs

pip install torch pyyaml tqdm wandb numpy scipy

wandb login  # or disable with: export WANDB_MODE=disabled

The uniform baseline model is designed to serve as a simple reference baseline. It applies a random compression strategy for the synthetic and real-world datasets. You can run this baseline using the following commands:

python benchmark/experiments/uniform_baseline_compression_test.py

It should complete in about a minute without any GPU-acceleration.

To run the graph sampling experiment using the uniform sampler, run the command:

python benchmark/experiments/uniform_baseline_graph_sampling.py

We've developed three CUDA-compatible probabilistic Knowledge Graph Embedding models: TransE, DistMult, and ComplEx. Run experiments using the commands below:

# syn-paths
python benchmark/experiments/probabilistic_kge_baselines.py --config benchmark/configs/syn-paths-[model].yaml

# syn-tipr
python experiments/train_baseline.py --config benchmark/configs/syn-tipr-[model].yaml

# syn-types
python benchmark/experiments/probabilistic_kge_baselines.py --config benchmark/configs/syn-types-[model].yaml

# wd-articles and wd-movies
python benchmark/experiments/probabilistic_kge_baselines.py --config benchmark/configs/wd-[dataset]-[model].yaml

Replace [model] with transe, complex, or distmult and [dataset] with the appropriate dataset name.

We have written test functions to check the graphs in the datasets against the list of rules. It can be run using:

python intelligraphs/data_validation/validate_data.py

If there are any errors in the data, it will raise a DataError exception and the error message will look similar to this:

intelligraphs.errors.custom_error.DataError: Violations found in a graph from the training dataset: 
        - Rule 6: An academic's tenure end year cannot be before its start year. The following violation(s) were found: (_time, start_year, 1996), (_time, end_year, 1994).

If you use IntelliGraphs in your research, please cite the following paper:

@article{thanapalasingam2023intelligraphs,
  title={IntelliGraphs: Datasets for Benchmarking Knowledge Graph Generation},
  author={Thanapalasingam, Thiviyan and van Krieken, Emile and Bloem, Peter and Groth, Paul},
  journal={arXiv preprint arXiv:2307.06698},
  year={2023}
}

If you encounter any bugs or have any feature requests, please file an issue here.

IntelliGraphs datasets and the python package is licensed under CC-BY License. See LICENSE for more information.

This package has been and developed and tested on MacOS and Linux operating systems. If you experience any problems on Windows or any other issues, please raise the issue on the project's GitHub repository.

Make sure to activate the virtual environment with the installation of the intelligraphs package.

To run the unit tests, install pytest:

pip install pytest or conda install pytest

pytest --version  # verify installation

Execute the units tests using:

pytest

If you would like to contribute code for a new feature or bug fix, here's how to get started:

First, set up your development environment:

git clone https://github.com/thiviyanT/IntelliGraphs.git
cd IntelliGraphs

python -m venv venv
source venv/bin/activate  # On Windows, use: venv\Scripts\activate

# Install development dependencies
pip install -e .

For submitting changes:

# Create a new branch from dev
git checkout dev
git checkout -b feature/your-feature-name

# Make your changes and commit
git add .
git commit -m "Description of your changes"

# Push to GitHub
git push -u origin feature/your-feature-name

To submit changes:

1. Ensure all tests pass by running pytest.
2. Update the README.md, if needed.
3. Create a pull request from your feature branch to the dev branch.
4. The CI pipeline will automatically run tests on your pull request.

Changes must pass all tests and be approved before they can be merged into the main branch. For questions or discussions, please open an issue on GitHub.