https://imgur.com/RPhHG5W.png

PANINIpy: Package of Algorithms for Nonparametric Inference with Networks in Python

PANINIpy: PANINIpy is a Python package providing a collection of nonparametric statistical inference methods for unsupervised learning with network data. Many of these methods are motivated by the Minimum Description Length (MDL) principle and aim to make network inference accessible and robust without the need for parameter tuning. Some benefits of using PANINIpy are listed below:

  • Easy Installation: PANINIpy can be easily installed through pip.

  • Pure Python Implementation: All modules are written in pure Python, ensuring compatibility with standard Python environments and minimal dependencies.

  • Parameter-Free: PANINIpy’s nonparametric algorithms allow users to extract meaningful summaries of network data without the need to manually adjust free parameters. This reduces potential biases from human aesthetic preferences and implicit algorithmic constraints, and makes the network inference process more user-friendly.

  • Statistical Robustness: The methods are designed to separate significant structural regularities from noise in network data, making them suitable for real-world datasets that may contain statistical fluctuations and measurement error.

  • Focus on Network Inference: Unlike broader network analysis packages, PANINIpy specifically focuses on nonparametric inference, filling a gap for specialized statistical analysis without the overhead of broader functionalities.

Statement of Need

Due to their discrete, relational, and heterogeneous nature, complex networks present new obstacles for statistical inference. Many inference objectives on networks are intrinsically combinatorial and produce complex summaries in the form of sets or partitions. These factors make scalability and interpretability of critical importance for practical algorithms, which are not often easily accommodated within learning frameworks that focus on continuous ordered data. There are also a number of ways uncertainty can be introduced in the collection of a network dataset, whether through measurement error, sampling bias, or fluctuations across experimental settings in longitudinal or cross-sectional studies. These factors underscore the importance of developing new principled, nonparametric methods for extracting structural and dynamical regularities in networks that do not rely on ad hoc parameter choices or heuristics, allowing them to be robust in the presence of noise. These methods should be tailored and optimized for specific inference settings (e.g. network population clustering or hub detection) for the best performance in practice.

PANINIpy is a flexible and easy-to-use collection of nonparametric statistical inference methods for unsupervised learning with network data. These methods are unified in their motivation from fundamental principles–currently, the Minimum Description Length (MDL) principle underlies all the methods in PANINIpy—and their lack of dependence on arbitrary parameter choices that can impose unwanted biases in inference results. PANINIpy is highly accessible for practitioners as its methods do not require the user to manually tune any input parameters and everything is written from scratch in pure Python to be optimized for each task of interest.

PANINIpy fills an important gap in the software space by focusing on nonparametric inference methods for tasks beyond community detection and network reconstruction, for which there are many well developed and maintained Python packages (including Graph-Tool, NetworkX, iGraph, and netrd among others). There are no existing Python packages allowing for the breadth of network inference problems tackled by PANINIpy, which provides methods for network hub identification, temporal and multilayer network aggregation, spatially contiguous regionalization, and network backboning among other methods. By providing a unified package with these methods, users can identify parsimonious summaries of their network data from multiple perspectives, all comparable on the absolute scale of data compression in bits (for the MDL-based methods). PANINIpy does not have the extensive dependency requirements of existing packages and tailors its data structures for each specific network inference problem for efficient algorithmic performance and easy maintenance. PANINIpy therefore provides an important complement to existing network analysis packages in Python such as NetworkX that focus on network metrics and network visualization (with some methods for community detection), as well as Graph-Tool and netrd which focus on community detection and network reconstruction respectively.

Citation

If you use PANINIpy in your research or projects, please cite it as follows:

BibTeX Format

For BibTeX users, include the following entry in the bibliography file:

@article{Kirkley2024,
    doi = {10.21105/joss.07312},
    url = {https://doi.org/10.21105/joss.07312},
    year = {2024},
    publisher = {The Open Journal},
    volume = {9},
    number = {103},
    pages = {7312},
    author = {Alec Kirkley and Baiyue He},
    title = {PANINIpy: Package of Algorithms for Nonparametric Inference with Networks In Python},
    journal = {Journal of Open Source Software}
}

APA Format

For APA users, cite as:

Kirkley, A., & He, B. (2024). PANINIpy: Package of Algorithms for Nonparametric Inference with Networks In Python. Journal of Open Source Software, 9(103), 7312, https://doi.org/10.21105/joss.07312

How to Cite

To reference PANINIpy, one can use either the BibTeX or APA format provided above. Please ensure that the citation includes the DOI and the link to the official article for clarity and accessibility.

For more details, visit the official article published in the Journal of Open Source Software.

Installation and Contributing

Methods

Attribution

The logo for this package was enhanced using Stable Diffusion model, an AI-based generative model created by Robin Rombach, Patrick Esser and contributors.

The model is released under the CreativeML Open RAIL-M License. For more details on the model and its licensing, refer to the following: