Skip to content

We should use mixture distributions where appropriate rather than categorical #21

@mmcdermott

Description

@mmcdermott

For all non-multinomial distributions, we should use pomegranate to fit mixture models of the appropriate components pending the data type (e.g., positive only, discrete, normal, etc.) to the data then serialize the mixture model parameters rather than anything derived from raw data, and sample from the mixture models during generation. This will help with #20 and allow us to capture richer dataset dynamics and better visualize dataset properties with more concise, meaningful configuration files.

See https://chatgpt.com/share/67c6267e-a498-800c-8f0b-00463bc44656 for a discussion with ChatGPT on different technical approaches.

Metadata

Metadata

Assignees

No one assigned

    Labels

    GenerationFor issues relating to the generation of synthetic dataInference of Generation ConfigsFor issues relating to the inference of data generation configs.enhancementNew feature or requestpriority:highHigh priority; should be shortlisted for upcoming release

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions