PPNP & APPNP Models

Description

This repository hosts the official implementations of the PPNP (Predict then Propagate) and APPNP (Approximate Personalized PageRank) models, as detailed in the ICLR 2019 paper "Predict then Propagate: Graph Neural Networks meet Personalized PageRank" by Gasteiger, Bojchevski, and Günnemann. The project offers both TensorFlow and PyTorch versions of these advanced graph neural network models.

The primary goal of this repository is to provide researchers and developers with accessible code to leverage these models for tasks involving graph-structured data. The implementations are designed to be straightforward to use, with clear examples and instructions for reproducing the results presented in the original paper. This facilitates further research and application of Personalized PageRank concepts within the GNN framework.

Key capabilities include the ability to perform graph classification and other node-level or graph-level predictions by effectively combining propagation mechanisms with predictive models. The models are particularly suited for scenarios where understanding node importance and influence within a network is crucial. The repository also includes utilities for data handling, such as converting NetworkX graphs into the project's SparseGraph format, enabling the use of custom datasets.

The target audience for this repository includes machine learning researchers, data scientists, and engineers working with graph neural networks, network analysis, and related fields. Developers looking to implement or experiment with state-of-the-art GNN architectures for tasks like link prediction, node classification, or recommendation systems will find this resource valuable.

The value proposition lies in providing a well-documented, open-source implementation of powerful graph neural network models that bridge the gap between traditional graph algorithms like Personalized PageRank and modern deep learning techniques. The availability of both TensorFlow and PyTorch versions enhances flexibility and adoption across different development environments.

PPNP & APPNP Models Highlights

• TensorFlow implementation of PPNP and APPNP models

• PyTorch implementation of PPNP and APPNP models

• Code for reproducing results from the ICLR 2019 paper

• Example notebooks for easy integration (simple_example_pytorch.ipynb, simple_example_tensorflow.ipynb)

• Utilities for converting NetworkX graphs to SparseGraph format

• Includes datasets like Cora-ML, Citeseer, and PubMed

• Supports graph-based machine learning tasks

• Based on the "Predict then Propagate: Graph Neural Networks meet Personalized PageRank" paper

• Open-source code available on GitHub

Getting Started with PPNP & APPNP Models

1. 1

   Access Model: Clone the GitHub repository to your local machine.

2. 2

   Set Up Environment: Install required packages using `pip install -r requirements.txt`.

3. 3

   Integrate via API: Import and utilize the PPNP or APPNP model classes within your Python scripts.

4. 4

   Run Examples: Execute provided Jupyter notebooks (`simple_example_pytorch.ipynb` or `simple_example_tensorflow.ipynb`) to see the models in action.

5. 5

   Reproduce Results: Utilize `reproduce_results.ipynb` to replicate experimental outcomes from the paper.

6. 6

   Data Handling: Use `networkx_to_sparsegraph` for custom dataset integration.

PPNP & APPNP Models's Use Cases

• Graph Classification
• Node Classification
• Link Prediction
• Recommendation Systems
• Network Analysis
• Attributed Graph Learning

FAQ from PPNP & APPNP Models

• What are the PPNP and APPNP models?

  PPNP and APPNP are Graph Neural Network models that combine propagation mechanisms with Personalized PageRank concepts, as detailed in the ICLR 2019 paper "Predict then Propagate: Graph Neural Networks meet Personalized PageRank."

• What programming languages and frameworks are supported?

  The repository provides implementations in both TensorFlow (versions >=1.6, <2.0) and PyTorch (versions >=1.5).

• How can I get started with the PPNP/APPNP models?

  The easiest way to start is by exploring the `simple_example_pytorch.ipynb` or `simple_example_tensorflow.ipynb` notebooks provided in the repository.

• Can I reproduce the results from the paper?

  Yes, the `reproduce_results.ipynb` notebook is included to demonstrate how to reproduce the results presented in the original paper.

• What are the dependencies for this repository?

  Key dependencies include numpy, scipy, TensorFlow (>=1.6,<2.0), and PyTorch (>=1.5). NetworkX is needed for certain data conversion methods.

• How do I install the required packages?

  You can install all requirements by running `pip install -r requirements.txt` in your environment.

• Can I use my own datasets with these models?

  Yes, you can use custom datasets. The `networkx_to_sparsegraph` method in `ppnp.data.io` can help convert NetworkX graphs to the required SparseGraph format.

• What datasets are included in the repository?

  The repository includes several datasets in the `data` folder, such as Cora-ML, Citeseer, and PubMed, along with references to their original sources.

• Who authored the PPNP and APPNP models?

  The models were proposed by Johannes Gasteiger (formerly Johannes Klicpera), Aleksandar Bojchevski, and Stephan Günnemann.

• Where was the paper published?

  The paper "Predict then Propagate: Graph Neural Networks meet Personalized PageRank" was published at ICLR 2019 (International Conference on Learning Representations).

• What is the license for this code?

  The code is released under the MIT license.

• How can I cite this work?

  You can cite the paper using the provided BibTeX entry: @inproceedings{gasteiger_predict_2019, title = {Predict then Propagate: Graph Neural Networks meet Personalized PageRank}, author = {Gasteiger, Johannes and Bojchevski, Aleksandar and G{"u}nnemann, Stephan}, booktitle={International Conference on Learning Representations (ICLR)}, year = {2019} }