OK, let's try to answer this better. I'm not going to illustrate or point to the right references. I just want to throw some pointers. Beside, I write about my vision which is not mainstream (and I believe it is not possible not to be personal about this: experts do not agree for the moment).
"In essence, Mathematical Optimization (MO) and Machine Learning (ML)
are different but complementary technologies."
Yes and no. A few years back, ML (or parts of ML) was considered as a sub field of MO... ML uses MO to optimize its predictions (for instance, to minimize the discrepancy between its predictions and the reality (supervised learning)). But yes, they are complementary to say the least as one (ML) predicts, the other (MO) optimizes. The frontier between both is blurry.
"Simply put – Mixed Integer Programming (MIP) answers questions that
ML cannot. Machine learning makes predictions while MIP makes
decisions. They call Artificial Intelligence (AI) the combination of
both technologies."
Exactly, except that some AI experts kicked MO out of AI. And that AI is much broader than that.
"More specifically :
ML predictions can determine the need to make a MO decision. ML
predictions can be used as MO decision constraints (when the data is
missing)."
Sorry, but I don't understand this. ML is able to find patterns in your data and reproduces up to a point the "past" (it can be used to innovate and optimize but it is not its strongest point: my view). If you want to be pro-active and make decisions based on optimization (and not observations) then you use MO. If you want to make decisions based on observations (and not optimization) you use ML. At Funartech, we combine both to take decisions. ML methods are data resources hungry, i.e. it is almost impossible to use ML if you do not have (clean) data. And most of the time, you need huge amounts of data.
"Other interesting examples are given here by FunArtech, a startup in
Montreal who claim to be the future of AI. For example, they claim
that "some research is conducted in order to predict on what variable
and how to branch in search trees.""
We do claim so because we have better practical and theoretical results. Yes we do.
In a very basic and scholar approach, you could divide the use of ML and MO in four categories:
the combination of ML and MO as two black boxes used separately. This has become mainstream in Québec (but not in Silicon Valley for instance) and other parts of the world. There are several examples on our website (https://www.funartech.com/approach/use-cases).
and 3. You can use one to improve the other. MO can be used to optimize and improve ML (this IS ML...) and it goes the other way too: ML can be used to improve MO algorithms. The use of ML to branch in a search tree is one such example.
we have found a new hybridization of both fields with new algorithms. This is an industrial secret for the moment.
"This is interesting and leads me to asking the following (broad, I
agree) questions :
In your personal experience, as an OR scientist, to what extent and
how have you used ML for MO ?"
I my personal experience, I mix both all the time.
"What would you recommend MO practitioners to learn in priority ?
General ML techniques ? Neural networks ? Data statistics ?"
I would recommend to first figure out what type of problems you want to solve and to dig the corresponding ML/MO parts needed to solve such problems. ML is not easy but as a MO practitioner you have a base knowledge that will help you understand the mathematics of ML better than most ML experts that don't have such a mathematical background. Even if the ML field is broad (huge!), the basic principles and ideas are "small". The level of mathematics required is the one of a first year student at university, so totally accessible.