There are many misconceptions about machine learning and artificial intelligence (AI). One prevalent misconception is that the terms are interchangeable. We’ve dispelled this notion in a previous blog. Machine learning is a subset of AI. Self-driving cars provide a good illustration of the differences between the two. Autonomous vehicles require hardware for sense-and-control, software to recognize situational details (traffic, lights, lane changes, etc.) and an “executive function” that acts on the input and drives the car. Machine learning provides the pattern recognition component, learning and predicting what might happen next. AI replaces or augments “intelligent” human vehicle control tasks. Machine learning uses historical data to create descriptive, predictive, or prescriptive models that users can run against current data, whereas AI uses these predictions to control one or more processes.
Often machine learning, AI or learning systems in general are bandied about as synonyms for “analytics.” The truth is, straight-up analytics – a graph or a plot of numbers, a calculation of sales over time, etc. – do not necessarily involve a learning system. Machine learning can produce analytics as an outcome, but not all analytics is machine learning.
There are other common misconceptions about machine learning that are important to debunk, especially for marketers looking to tap into the power of machine learning to help deliver hyper-personalized customer experiences at scale. Machine learning can be more useful, and provide better insight than run-of-the-mill analytics.
Machine Learning, a Trusted Ally
One misconception is that, for whatever reason, machine learning systems in general cannot be trusted. This notion seems to be hardwired; unless we see it with our own two eyes, we tilt toward disbelief. As my colleague George Corugedo points out, it’s human nature to think a humanized customer experience must, by definition, include directed human involvement.
Another common mindset is that something that has worked before must therefore work again – the “we’ve always done it this way” syndrome. It can be difficult to break out of this rut; we’re conditioned to think the techniques we’ve learned must always deliver the right answers. In a highly dynamic world, however, mistaken assumptions may often lead to right answers to the wrong sets of problems.
To minimize the number of assumptions, marketers must accept that the world moves faster than we can keep up doing things manually. The world is dynamic – markets, consumer behavior and interactions change all the time, and some of this is reflected in data that changes by the day, hour and second. A great model may go stale quickly. Marketers, therefore, need to be open to alternative technologies and to understand that these offer the ability to keep pace with dynamic environments.
One way to alleviate the mistrust is for marketers to understand that machine learning does not need to automate everything in a system; there can be parts that are manually controlled, or that require some level of manual intervention such as approving results. A recognition that machine learning is not an all-or-nothing proposition will help debunk the similar misconception, which is that automation in general will end up replacing humans (read: marketers and data scientists.) There is a perceived job security factor in play. Or rather, job insecurity. Alleviating this involves adopting a new mindset that machine learning is, at its core, about making things more efficient – outcomes, processes and, yes, people. Testing and applying machine learning in the context of the business will always need smart people.
With the understanding that machine learning is a collaborative partner rather than a system designed to eliminate marketing jobs, marketers will more fully appreciate the learning system as an enabler of more powerful human-centric experiences.
Real Problems, Real Solutions
Another common misconception about machine learning is that it is a magic bullet that solves all problems perfectly, just by throwing enough data at it. Some of this misconception can be attributed to the buzzword factor. Like “Big Data” of a few years ago, machine learning and AI suddenly became something every business had to have – even though few understood why.
To dispel this notion, it’s important to understand not only what problems machine learning is good at solving, but also what areas are (currently) not a good fit. Underlying this thinking is the need for businesses to be crystal clear about a use case. Deciding in advance on a judicious application will help debunk the notion that machine learning, provided enough data, will solve any problem.
Just like human-driven solutions, machine learning also must work within the constraints put on the business by regulatory agencies, data privacy laws, overarching corporate objectives and even physical constraints. The saying that “to a hammer everything looks like a nail” is analogous to how some marketers view machine learning – as a tool that solves every problem. The reality is, just as a toolbox is filled with different gadgets, the key is to use the right tool (or tools) to solve the problem at hand.
Freed from the Coding Burden
Another myth about applying machine learning is that it must require heavy coding by data scientists. But because the world is dynamic, with data and customer journeys changing rapidly, is precisely why machine learning needs to be an easy-to-configure, automatable process – from data inputs through to model building, deployment and optimization. Automating these processes without the burden of coding is key to keeping pace with business cadence.
Marketers that harbor a belief that ML requires coding then think they will either need to hire more people, learn how to code, or send data off to a third-party vendor to build a model. It’s even worse if the third-party vendor requires a licensing agreement to use the model. Having to repeat this tedious manual process every time there’s new data is very inefficient. Automation ensures the model comes back in time to still be relevant to the problem you’re trying to solve.
Predictive Modeling vs. Business Optimization
Another misconception, somewhat in line with the hammer-nail analogy, is that many marketers fail to distinguish between using machine learning for predictive modeling and for optimizing a business process. Optimization is an integral part of any successful business operation, and is a key component of learning systems. Think of optimization as ‘knob-tuning’ – finding the best set of configurations, parameters, process sequences, etc. to meet one or more goals. I.e., basically a search through a set of ‘what-if’ scenarios, subject to applicable constraints (resource limitations). Obviously, this involves prediction – this is where machine learning modeling comes into play. The optimization part takes these predictions, uses them to evaluate how well things would be, then tunes the salient parameters to get the ‘best’ expected outcomes.
For example, in the world of customer engagement segmentation models are often used to predict which offer or recommendation will be most relevant to a customer. Once predictions are available, they might be used to evaluate profitability. The optimization piece can use these predictions toward tuning the parameters with a goal of producing a business outcome – such as increasing customer lifetime value or reducing attrition.
Airplane scheduling is another good example. If the business process is to optimize gate turnover to increase profit, machine learning will be used to model a host of processes, from predicting time spent onboarding and offloading, to cleaning, fuel consumption, routes, weather, maintenance, etc. With the ultimate goal to increase profit by having more planes takeoff from one or more locations, optimization will find the best setting(s) for each variable to produce the desired outcome.
This misunderstanding, as it were, brings us full circle to the first point mentioned – the notion that machine learning is a synonym for analytics. In the airline example, analytics are the statistics – jet fuel consumption from Point A to Point B, average onboarding time for a 737, an Airbus 340, etc. Machine learning could be used to predict and generate these statistics. Artificial intelligence might be used to control the flow of people, or direct support services. Optimization would decide the ‘best’ allocation of resources, where to locate them, and what schedules to fly, all with the goal of optimizing profit. The combination of these technologies create the opportunities for business to stay ahead of the game.
With a finer appreciation for what machine learning is – and what it is not – marketers will sidestep the common misconceptions, and begin to appreciate machine learning as an indispensable enablement tool for the delivery of hyper-personalized experiences at scale.