Is it ok (ethically speaking) to rename features of a dataset for publishing?

Question

I work in the machine learning field where I deal with datasets provided by an industrial partner, and one concern of the project is the confidentiality of the data.

My team is working on a fault detection system using those datasets which includes features or columns, with names as Motor_sectionA_speed, Motor_sectionB_torque, Valve_sectionC_pressure, etc. which are subparts of a big system, and if the context is known, they could be traceable to details of our partner machinery and operation.

For publishing some results two options have appeared regarding naming those features:

• Name features as Feature A, Feature B, Feature C, etc: I have seen this for synthetic datasets, where the focus is to highlight the algorithm where the importance of the feature is in its nature (distribution, range, etc) not its name or meaning.
• Name them as Motor_1, Motor_2, Valve_1: One person stated that from training she/he had, the previous option could be unethical because the meaning of the variables is lost and might be misleading. Instead, names can only be simplified as Motor_1, Motor_2, Valve_3, etc.

Is it the first option considered unethical in all cases? or is this a "depends"/gray zone matter?

Answer 1

There is nothing unethical about renaming features unless you do so with the purpose to mislead or fabricate results. It makes perfect sense to rename features to keep a company secret.

However, a publication stating (for example): "I have some features and use them to predict manufacturing errors with 99% accuracy" has no value whatsoever: for all the reader knows you could be using the phase of the moon, dice rolls, and the position of tea leaves in your cup. The reader would be unable to tell what is going on, and unable to verify, reproduce, or use your results. This defeats the point of publishing and the paper would probably (and hopefully) not be accepted.

If you wish to publish something else, such as a new algorithm, it may be possible to use an entirely different dataset. You can still mention the actual application you worked on and simply explain that these data cannot be shared.

Also, you may be underestimating the ingenuity of others and the power of data. Are you really sure that renaming the features is enough to prevent others from deducing what is going on?