4

I am an Industrial Ph.D. and the company involved in my Ph.D. provides data for my research.

I signed an NDA with the company since the data comes from proprietary source I cannot disclose.

The main problem is that when I write articles I cannot disclose anything about this data (I am in the Machine Learning field), and this decreases the validity and reproducibility of my results.

It is really frustrating. Any Idea on how to tackle this problem?

3

There are a few solutions:

  • Use a different dataset for your paper. For example, if you design a new data clustering algorithm but cannot use the company data to evaluate your algorithm, you may find alternative datasets that could be OK to use just for the purpose of writing the papers. I have done this for some industrial project.
  • Some companies may accept to release anonymized or transformed data. You may try to discuss this with the company to see if this is a possibility.
  • Some journals and conferences may request that you share your data and code publicly. But some other may not be so strict about this. Choosing appropriately conferences/journals may help.
  • You may add some text in your paper to explain that data cannot be made available due to a NDA. Reviewers may be satisfied by this explanation or may not be, but it can be good to do that.
  • You may consider submiting in conferences where there is an industry track. Your paper may be more well-suited for this than for the regular track.

I would also add that doing academia/industry collaborations is great but not all companies have the same level of openness toward publishing papers. It is generally better to make things clear about how paper can be published as early as possible in a colloration. When looking for a new industrial collaboration, I will always talk with the company first about whether or not I can publish papers and what are the requirements to make things clear. Some company will ask you and even push you to publish papers, while others are against it. I even talked with a company that allowed to publish a paper but did not allow to publish the developed algorithm... (which makes no sense...as you then cannot write paper with just an introduction/related work and then a result section without describing the algorithm/method).

2

I sometimes used to supervise undergraduate dissertations through the process of agreeing an NDA with an industrial partner company. The university's internal IP team would usually carry out the initial drafting of the NDA. The university IP team's draft would invariably subject students and academic staff to restrictions onerous enough to be damaging to the dissertation process and to the student's subsequent graduate job search. In every case (*), it turned out that the industrial partner company did not particularly want those onerous restrictions, and was happy to agree to my request that they be dropped from the final version of the text of the NDA. In your case, the NDA is already signed, so that ship has sailed. However, my experience leads me to believe that, if you and your supervisor ask the company to release you from the requirements of the NDA for a particular publication, there's every chance they'll say "yes". (But make sure to get that "yes" in writing, and from an officer of the company who has legitimate authority to say it.)

(*) At least, in every case for which I was supervisor. Some colleagues in my group supervised projects for which it was otherwise.

1

I am in a bit of a dilemma about whether this deserves a separate answer, as it is similar to the first suggestion in Phil's answer, but I do think it is slightly different.

In any research field, therefore also in Computer Science and Machine Learning, we should strive to offer theories and approaches which generalise well to (m)any conditions and data. So I would suggest validating your results on more than one dataset:

  • the company-provided one as this was the motivation for developing your approaches.

    You could try asking for permission to describe the characteristics of the data, and maybe include a (single, or a handful of) representative example(s). This might often be possible, as one single sample has little value and could be anonymised if needed.

  • publicly available data, to facilitate easier comparison with other approaches.

    About the comparison: it not only allows future works to compare with yours, but also allows your work to be presented in comparison to state-of-the-art, strengthening your contribution. As an additional perk, this demonstrates that your approach is general and not dataset specific.

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.