Theoretical limitations and embeddings

[thiswillbeyourgithub]

Hi,

I'm a psychiatry resident and computer scientist interested in using open-cesp for synthetic dataset generation.

I was wondering the theoretical limitations of your sampling technique regarding different data types, particularly high-dimensional data like text embeddings?

Also, I'm concerned about potential privacy leakage with what I call "rich" data formats, specifically text embeddings. Recent work (https://arxiv.org/html/2505.12540v2) has demonstrated that text embeddings can be reverse-engineered back to original text without even knowing the embedding model used.

Given that your technique preserves statistical characteristics while anonymizing data, I'm wondering:

1. Are there specific data types that shouldn't be cloned with this approach? If unsure, are there established methods to check how well the sampling did and how de anonymised it is?
2. Have you evaluated it for embedding-based or high-dimensional continuous features?
3. Are there recommended preprocessing steps for such rich data?
4. Are there types of data that could be cloned but would be computationally intractable? Any useful ballpark figures to give me an idea of the scales here?

Thanks for this interesting project!

[sauravk58]

@thiswillbeyourgithub assign me

[thiswillbeyourgithub]

@sauravk58 I don't think I have the appropriate rights to assign users to issues. Have I?

[sdv-team]

Hi @sauravk58, it’s great to see your interest in the SDV ecosystem. The SDV software (and its related libraries) is owned and maintained by DataCebo. It is available under the Business Source License for you to browse.

We support a large set of users with enterprise-specific intricacies and reliability needs. This has required us to be deliberate about setting the roadmap for SDV libraries. As a result, we are unable to prioritize reviewing and accepting external pull requests. As a policy, we will not be able to accept external contributions.

If you haven't already, we would greatly appreciate it if you could file an issue instead with the overall description of your problem. We can determine whether it’s aligned with our framework. Once discussed, our team typically resolves smaller issues within a few release cycles. For more information please see our Contributing Page. We appreciate your understanding.