Machine Learning and AI — What’s ‘Real,’ What’s Required

Quite a while ago, big data had its beginning within the realm of academic research. Niche businesses then began implementing big data after recognizing its usefulness. Next, massive companies (like Google) started commercializing big data — these large companies then spawned smaller companies. Those companies are now getting big.

This all makes for a lot of noise in the marketplace.

Today, we hear folks without applied mathematics or computer science backgrounds talking big data, algorithms and artificial intelligence (AI) at cocktail parties. The fluency has grown rather quickly: A CMO I’ve known for years used to wince when we talked analytics, but now she enthusiastically discusses her firm’s AI initiatives. She’s not running marketing at Google or IBM Watson, either — she sells clothing online.

While we’re likely in one of the most amazing periods in history to be in business, it does not come without its challenges. These days, you have to sift through all of the clutter when it comes to innovations in the marketing space.

Let’s see if we can simplify what the data pundits are tweeting and discern where the value really is.

Machine Learning

Machine learning (ML) occurs through networks of algorithms.

First, the good news: ML really works.

As we’ve discussed in “Marketing Machines — Possible or Pipedream?” ML is used to ingest large amounts of data and identify patterns in that data. The machine “learns” by ingesting, transforming and then conditioning a learning algorithm with your dataset.

ML will find the statistical relationships (models) between your various data points to articulate how efficiently your business is running. By calculating the best potential models, it can also show you what improvements you can make. ML can deduce your most profitable business targets. It can tell you who is likely to buy shoes priced over $800, or which production line is most likely to break down in the wintertime.

But ML Isn’t Foolproof

Machine Learning can surely help us find structure and patterns in data through statistics and the power of cloud computing. Amazon’s ML cloud computing capability, for example, isn’t specific to any domain and arguably works with any inputs. It will consistently output a result or target. Yet that very flexibility is where ML can prove risky:

“If you can dump anything into an ML process, and have it come up with an answer, you’d be wise to be wary of that answer.”

ML techniques all require you provide it with a “universe”. This universe consists of all the likely permutations representative of your purpose. If your conditioning data is skewed heavily to sneakers under $75, it will prove very hard to predict what customers are likely to buy $800 shoes.

This may sound like an unfair example, but consider the marketers who are out to break into the higher-end sales but only have data from their pre-existing customers. If skewed interpretations were applied to new-customer marketing (and they can be), your returns could be even worse than without any ML interference. The fact is, there are far more experiments where ML doesn’t produce a valuable outcome than those that do. But as technology and big data are refined over time, better results will be achieved across the board.

Analytics and model-building are highly iterative processes. If an ML process is focused on only a particular niche, the likelihood of getting better results sooner is higher — but still iterative. Despite its current limits, AI offers a deeper and more layered method of applying iterative math to break down large data questions than raw manpower.

Google’s AlphaGo AI beat champion Lee Sedol in a tournament of Go by exponentiating component questions, covering as many bases as it could. While AlphaGo works similarly in many ways to the human mind in this way, it did also have the advantage of iteratively playing against itself thousands of times.
Humans can’t do that.

The Bottom Line: Good Data In, Good Comes Out

Whether Google’s AlphaGo, Amazon’s ML tools or your home-grown mashup, the quality of the data that goes into ML is the largest factor you can control in creating value with systems-driven optimization.

In an age where many organizations have siloed data or cumbersome messes, along with marketing organizations that don’t even have a reliable marketing operations database, this is no small challenge. Getting your data centralized, organized and accessible is a requisite first step. Get that right, and there may be opportunities ahead to drive value up.