Hello Hussien,

Assuming I understand what you're after, the cosine similarity (akin to correlation among vectors in space) is unlikely to help you accurately place a previously untested method relative to others with respect to level. Association is a check on score patterns, not levels.

Correlation/association can be high, while levels are quite disparate; as well, correlation/association can be low, while levels are quite similar.

Good luck with your work.

You can look into the K-NN algorithm in this context.

As far as I understand your problem, you first need a mathematical relation between the instructions and rank. For instance, Rank x should correspond to some instruction value as y and vice versa; it means you require a mathematical function.

So there are various methods/tools to find a suitable (as accurate as you want) to find mathematical function based on given discrete values like curve fitting methods or the use of ML.

Further, Once you obtain the mathematical function, run your code a few times, and you will get a set for various combinations of (instruction, rank). These set values will work as the feedback for your derived function. Make changes based on the feedback, and you will get a much more accurate function.

I hope you are looking for the same.