Machine Learning And AI – Don’t Buy Hype, Buy Solutions

Machine Learning And AI – Don’t Buy Hype, Buy Solutions

by Pat Stroh , Op-Ed Contributor, August 17, 2017

To say machine learning is over-hyped is an understatement. And it would be wise for marketers to not get too dazzled by this shiny new tool. Here’s why in one sentence: new advances in machine learning are often unnecessary and undesirable if the objective is to gain business advantage from analytics-driven marketing. “Whoa, whoa, whoa” some might say. Let me explain.

First, let’s settle on a definition of machine learning. My own definition is: Machine learning (ML). The use of (mathematical / statistical) algorithms in the context of very large, fast moving, and highly complex data sources (i.e., “big data”).

It’s a stretch to use the term when referring to small, slow moving and simple data sources. Artificial intelligence (AI) is also an overarching term used to describe machine learning. Without a doubt, new developments are successfully addressing areas where there are vast volumes of unstructured, fast moving and noisy data. But there’s one more feature of ML/AI that warrants attention and that’s the degree of transparency. How do the algorithms get from the inputs to the outputs?

Machine Learning And AI - Don't Buy Hype, Buy Solutions |

ML/AI are not very transparent. By that, I mean it’s difficult to explain or understand the precise manner that produces a prediction or output. The term “black box” is often used. Here, ML/AI is largely judged by assessing the outputs alone, as opposed to the specific steps involved in producing the output. In more technical terms, the “error of the output” is of more importance than any mediating steps. Hence, we get bots that learn to be racist. (How did that happen?) Make healthcare decisions that doctors don’t understand. And serve content (ads) that just seem odd. For marketers, this is a decidedly mixed blessing.

In contrast to advanced ML/AI, many “conventional” statistical algorithms are relatively transparent. It’s much easier to understand the algorithm, and the specific reasons for an output. The reason for customer imminent churn is a (weighted) combination of tenure, servicing issues, and price changes. Product A is recommended based on the customer’s segment, purchase history, and social buzz. The strong performance of digital paid search and email caused sales, as opposed to direct mail. The reasons for the output, here, are integral to marketing decisions.

Therefore, as we all move toward a more data / analytics-focused marketing ecosystem as we should be doing it’s important for marketers to avoid becoming dazzled by the new ML/AI tools. They are powerful tools in certain contests … big, fast, and complex data sources (such as language comprehension, image processing, IoT/device analytics, fraud detection). However, conventional analytics in many cases are far more useful when dealing with relatively small, slow moving, and structured (simple) data sources.

Analytics is a powerful tool, and it can deliver business and marketing success in so many ways. Common problems such as segmenting customers, decreasing customer attrition, targeting the most responsive prospects, optimizing price points and measuring marketing ROI, etc. can all be solved through analytics … conventional small, slow, structured data analytics. We should not unnecessarily overcomplicate the need for advanced ML/AI algorithms.

In conclusion, if assessing the uses of analytics whether conventional analytics or ML/AI I’d recommend focusing on the data situation in combination with the solution you desire. That process will lead to much greater success, and less disappointment, than simply buying the top shelf tool. Search Marketing Daily