Novelty, importance and relevance
The main criteria for publishable research in ML as in pretty much every other field is novelty, importance and relevance.
Is your proposed solution or technique novel, and can you demonstrate that? You'd need to provide an overview of relevant literature and alternative solutions, and show how what you're proposing is new and different (at least slightly) from everything that has ever been published on that topic.
Is your proposed solution or technique important and interesting for others? If there's a bunch of published research on similar problems, you'd need to demonstrate why your paper is either better in some aspect or would be otherwise interesting for all the people working on these other problems; if not, then you'd need to show why that problem is important and why others would care about it - if it's a very unique task for e.g. your employer and noone else, then it's generally publishable if and only if that solution would be interesting and inspiring for people solving similar problems and that there actually are people solving similar problems.
And last but not least, your problem and solution needs to be relevant to the specialty of the venue where you want to publish it. A Nobel-prize worthy chemistry experiment won't (usually) be relevant enough to be published a machine learning conference; and the same applies for narrower niches within ML - certain venues are looking for certain types of papers only and would reject other topics even if the papers are really good.