How Machine Learning will solve the business problems of the future

This afternoon I met the losing end of a prolonged and impassioned negotiation with Alexa. I was in search of a precise word combo. Until I landed it, the living room light would remain off. Alexa meanwhile, wouldn’t give a bloody inch. Then came the realisation that I was standing next to said light switch. Refusing the indignity, I retreated upstairs. 

It’s February 2017. We live in exciting times. Some are optimistic that machine learning (ML) will emancipate mankind, whilst others draft policy initiatives at the highest level to prevent an accidental robotic military coup. Both eventualities are silly. Instead, ML will guide artificially intelligent drones to deliver you sweeties and your car will drive you to Sainsbury’s. It’s going to be really cool and surprising. Hell, even Alexa may be gracious enough to work the lights as promised.  

In these exciting times, the aspirations of Data Scientists betray no fear. Deep Learning is haute couture. You’d be forgiven for thinking that fancy algorithms well before they improve your lifestyle, are going to take your job or industry the way of the chocolate covered wagon wheel. Disruptive technologies do so every day. ML does not. Let’s not conflate the two. Heavy duty algorithms are not going to save the planet nor wage war. No one will be required to rise up against the machines. Sorry about that, if you are into that kind of thing. Rather, as we test the boundaries of ML, we are haunted on a daily basis by the limitations of algorithms.  

Despite this, the majority of Data Science literature focusses on reinvention, worshipping the slightest variation on algorithms already decades old. Data Science is also ridiculed, portrayed as a decision tree to help you select the most appropriate algorithm for the task. 

Throw your data at the right algorithm, tune some parameters and cha-ching! Such is the fantasy of the Data Science ‘Bootcamps’ and Marketing Managers. 

Behold the top of Kaggle’s leader boards. Here, the most consistently successful Data Scientists, titans of ML, pay increasingly more attention to the search for, understanding and manipulation of key predictors than they do finessing a model. Applied on real world datasets, disparate algorithms perform within decimal places of each other. Once deployed, this negligible difference is inconsequential. Algorithms are not important. Data are. 

The correlations we find therein must infer casual relationships otherwise they are just trouble. Despite the hype, this requirement of causality remains as inescapable to ML as it is to traditional analytical approaches, equally subject to GIGO. We’ll search high and low to isolate causal relationships yet so fierce is ML at hunting, it is blind to irrelevance and absurdity - so it often overreaches. We too, are as easily fooled by randomness as we are eager to make a good job of it. Datasets are everywhere. Useful data are not. 

Even when we have captured intuitive and expressive relationships, we pretend that they are static. Relationships are dynamic; markets evolve, viruses mutate, sounds and appearances are redesigned, climates shift slowly and human behaviour switches rapidly and irreconcilably. Laid bare, even the most sophisticated algorithm is a rule based equation. Shackled by limited scope and a short lifespan, it estimates tomorrow with a reasonable degree of probability based on what happened yesterday.  

In this regard, ML has more in common with the crazy lady on Brighton pier – it still only appears to foretell the future.  

More simply put, would you, dear reader, trust your worldly fortune to high performing algorithms? Of course not. No one is willing to swap out common sense and analysis when their own capital is on the line. That’s because the true risks are unknown and the costs of a Black Swan are too dear. e.g. 2016  

For these reasons, ML suffers a lack of universal relevance and while many try, only the smallest minority of investors see a return from ML. Success is apparent where there are enough labelled data, with actionable, stable relationships and when estimations instruct how the core business operates. Few business questions resemble this – it is a bitter pill for aspiring Data Scientists to swallow given how especially satisfying ML is to wheel out. It has sparked our imagination and its potential has egged us on. For a Data Scientist in February 2017, few things in life surpass the joy of doing a good job with ML. We celebrate both its sophistication and beautiful simplicity. But it is also by its very nature, awesomely lazy and now comes exquisitely wrapped up in 4 lines of copy and paste code.

ML is hurtling down the track toward becoming a cheap commodity.

It may take another five years to get there but get there it will, for it is no substitute for real business intelligence. Even at the very height of its powers, ML can supplement. It cannot replace.  

When questions demand exploration, interpretation, expertise and insight, the solution is never ever ML. Come the end of this big data revolution, we will perceive ML problems unworthy of curiosity. You can take that to the bank. So how does machine learning answer the business questions of the future? Sorry, I don’t know what you mean. I don’t know a playlist called Living Room Lite. Did you mean Lounge? Here is some smooth jazz.