In Data Science for Managers | Data Science Dojo, I teach a module on the importance of data. In this module is a slide titled “Machine Learning is Overrated”… not exactly the slide you would expect from a data science training company (or the data scientist delivering it!). Is ML overrated? I’ll present you three arguments for and against this statement and you can decide.
Yes, machine learning is overrated
Here are three reasons why ML is overrated.
Data is more important than the ML algorithm
As you might expect in a module on the importance of data, we emphasize that data is a supremely important topic for all analytics scenarios. My goal when doing ML is to find the minimum set of high quality data that gives me a reasonably high level of accuracy. I’ll sacrifice some accuracy for the explainability benefits of relatively few features and a simple algorithm. My takeaways documented in Using ML for the March Madness tourney – Lake Data Insights provide a real world example of this.
The step from descriptive analytics to predictive analytics is a big one
ML fits into the predictive category of analytics as it uses historical information to predict future states. It’s more complex than descriptive analytics which provide insights about history through stats and graphs. It’s less complex than prescriptive analytics which use predictions to provide recommended actions to the user.
An underappreciated cost of creating a robust ML solution is the effort required to enable and maintain it in production. The recent growth in MLOps shows how ML is maturing and the importance of operationalizing the solution. This is a big step that prevents successful analytics projects if not properly addressed.
There are a couple of alternatives that need to be considered before jumping to ML. First, a readily available visualization in the hands of an experienced user may give you the majority of the benefits of an ML solution. It could be even better than an ML solution that is a black box.
Second, a less costly approach to predictions could be hand coded heuristics. A domain expert could provide the logic needed for the heuristic and it likely will be cheaper in the long term than ML.
Too much hype
Being overrated is subjective and factors in both the capability provided AND the praise heaped upon it. Something can be very valuable yet still overrated. For example, an athlete can be both overrated and a valuable part of a team.
Many companies want to generate demand for AI capabilities. If you listen to them, you can be convinced that everyone is reaping huge benefits from their AI deployments. And you might also be convinced that anyone can do, that data science has been democratized so that you don’t need to have expertise in the field.
In reality, McKinsey’s State of AI report (Global survey: The state of AI in 2020 | McKinsey) indicates that only a few high performing companies have outsized returns for their AI projects. That same report shows that high performing AI companies are 2 to 3 times more likely to have continuous leaning programs for AI and resources available with data expertise.
No, machine learning is NOT overrated
Here are three reasons why ML is NOT overrated.
Where needed, ML can have huge impact
ML can have a huge impact in the cases where the business problem is large enough and the use of historical data for predictions is valuable enough. The targeted user base may not have enough domain knowledge to properly leverage descriptive analytics. The amount of data required to process is more than a human can handle. There are no simple heuristics that can be hand coded and maintained.
Amazon and Netflix can’t support their recommendation engines with humans. Predictive maintenance scenarios benefit from understanding all failure modes and not just the ones that one (or very few) technicians have seen. Matching host and guest preferences at scale is needed for AirBnB to succeed. And there are many more examples like this.
ML is a stepping stone to even larger ROI
While a single ML solution may not have huge ROI, bringing multiple ML capabilities together into a ‘system of intelligence’ may be a different story. A set of ML solutions in a single area may prove that the sum is greater than the parts and be a difference maker in your business.
The first ML solutions also build the human muscle and infrastructure capabilities to solve more ML based problems. You’re going to eventually need ML capabilities if you don’t want to be left behind, so you might as well jump on board now.
ML vendors are making significant strides
Whether we think that ML is overhyped or not, it’s clear that the vendors are making significant strides. You can take advantage of more ML capabilities implicitly when they are baked into the product. Canned ML capabilities like those provided by Azure Cognitive Services make it easy to integrate ML functions into your processes. Auto ML (Tools of the trade: AutoML – Lake Data Insights) reduces the expertise required to train a model.
These are just a few examples of strides that the vendors are making. And they will continue to improve over time to make ML capabilities more accessible to the masses.
Three arguments supporting the opinion that ML is overrated and three arguments against… which side do you land on???
Picture details: 0/9/2020, Canon PowerShot G3 X, f/5.6, 1/250 s, ISO-640