If you’ve been in the data or computational science field as a researcher or a professional in the last few years, you’ve probably come across someone talking about machine learning. You’ve probably come across an employer or a potential collaborator who was so excited about the potential of “machine learning” that they somehow managed to spit on you in their (pleasantly) ignorant but very cute excitement about the subject. (For those with a spitting problem, you should check out this piece about chronic spitting problems from 2005).
Of course, where hype and ignorance goes so does a fringe academic industry in hopes of capitalizing misunderstood and seldom used technical methods or technology trends, then the mainstream – afraid of falling behind – starts creating entire courses around the subject, then whole academic degrees, and before you know it everyone wants that amazing technical expert who knows all about machine learning, but have no idea what it is or what it’s used for. To add insult to injury those experts will charge outrageous fees for selling the hype along with everyone else.
Don’t get me wrong – this is not a piece about how useless machine learning is. On the contrary, machine learning is a useful method of analysis and prediction that has a rich and well-developed line of inquiry in academic literature, and has significant infrastructure in industry with some superb applications. However, like any other technical method, e.g. social network analysis, statistical analysis, genetic algorithms, systems dynamics modeling and so many others, it has specific uses where it shines and others where it does not make sense at all.
I don’t think it would benefit you if I go through the history or the details of the collective methods that amount to what we call machine learning today – only to say that the current machine learning developed in academia and used by industry amounts to nothing more than some statistical methods (regressions mostly), some clustering methods (which fall under classification), and some network analysis (like neural and Bayesian networks). Then there is a number of methods that focus on “learning”: decision tree learning, association rule learning, deep learning, reinforcement learning and so many more. It seems our friends of the machine learning community like to use the word learning (just kidding you guys) – which to me all of which just seems to be some combination or variation of the above methods – though I’m sure each sub-method has its own intricacies.
For some of us computational social scientists, it seems that so much of what we call machine learning today is just statistical and categorical analysis with some input-output matching to boot – valuable collection of technical methods and theory, sure – but is it worth the hype? any more than say the hype about “big data”??
The truth is that machine learning is an awesome tool for pattern recognition when you don’t know or have any expectation of what the pattern could possibly be, but the moment you know what the pattern could be, it lags behind in performance compared to basically ANY other specialized method where there is some closed form solution or some structured process or method of analysis. For example, say you are trying to identify who traditionally could be the central actor in a network – yes I’m sure, based on interactions, some attributes, and the network’s structure you can provide a machine learning process that will learn who the central actors tend to be – but if you have a closed form or well-developed solution like the ones we know network science provides us, then you can identify central actors much more efficiently, more precisely, and with a lot less computational resources than the machine learning process you put in place. That’s because you have already discovered the necessary relationships and forces that govern that particular relationship. This fact is also why machine learning is powerful in the analysis of unstructured data sets, and high-speed streams where there is no clear structure, relationship, or the time to find what the structure or relationship is to begin with.
The problem is, with all the hype, so many non-technical experts want to be able to say that they’re using this cool and new machine learning algorithm to discover or analyze XY and Z rather than make the effort and dedicate the resources to finding a permanent solution form, and I think that it will eventually hurt the usefulness of machine learning in industry, and potentially in the academic research arena as well – probably in the form of a lot of pseudo-science which will seem reliable but ultimately will be disproven with time.
So if you are an industry leader, or if you’re someone considering dedicating time and resources to machine learning, consider first WHY you need to use this powerful tool and whether it makes sense for your professional and academic goals or aspirations. This advice really goes with any scientific tool really, but right now it’s paramount we heed this advice in the machine learning arena.