Publishing a new dataset, based on existing

Question

Is it ethical and academically (or otherwise) acceptable to publish a new dataset, derived from an existing, well cited dataset?

The only novelty in this case are:

1. Newly computed features (based on published work), not currently part of the dataset
2. Data re-organisation, not existing previously, that aims to standardize partitioning of the dataset

The dataset in question pertains to Machine Learning.

(Related question at: Law Stackexchange)

Update: The only reason I wish to publish my version is to provide an option of common grounds for paper-methodology comparison. Currently, every author prepares his/her own desired subset and publishes the ML task results. This makes it extremely difficult to compare the papers or sometimes even simple reproducibility.

My proposed subset will be a standardized subset that every one can use directly and will help in comparison across papers.

Answer 1

As long as you cite the originals you avoid plagiarism. Citation removes the possible charge of plagiarism (unlikely in any case, here).

Just be clear about the origins and what you have done with it.

Also ask whether the license is even appropriate. In other words, does the dataset have enough "creative content" that it is copyrightable? Data per se can't be copyrighted. Only creative works can be. If it can't be copyrighted, then it can't be licensed. I don't have enough information here to help you make that judgement. But it would probably be good to make a conservative judgement.

But, you can also ask the authors for a more permissive license for your use if needed. They can do that for a copyrighted work.

The other question, of course, is how you intend to publish it and whether any publisher would be willing to work with you and consider your contribution worth publishing. This, of course, revolves around the issue of whether you have any creative content in the dataset.

Answer 2

I would suggest to just make the code available to generate the new features, and reference the data set you've taken from as well as the previously published methods that your code uses, rather than distributing the output.

The code is probably more broadly useful because it could possibly be used with other data, and if new versions of that data set come along your modification will be obsolete whereas your code will not (assuming the structure of the data does not change). If the combination of approaches you've used are themselves fairly novel, it may even be publishable as a paper as well, particularly if you have some results to show why doing these steps is useful in some way.

You could probably find a lawyer to argue that somehow you're allowed to do what you are proposing to do, but I don't see why. Even if technically legally allowed because of issues with considering data IP of any sort, it just feels a bit wrong to take someone's work and make it newly available with minor modification.