When it comes to new marketing technology obsessions, it’s easy to see why the likes of artificial intelligence (AI), cognitive computing and machine learning generate such hype.

SwissCognitiveThese evocative-sounding technologies conjure images of a brave new world where marketers can sit back, relax and let the machine do real-time, personalized, one-to-one marketing with effortless ease.

 

‘Intelligent’ Tech Needs Trustworthy Data

Few of the articles about those technologies ever seem to focus on the need for accurate, clean and trustworthy data to fuel such awesome capabilities. As close as we feel we are to some revolutionary future of AI-based marketing, many brands still do not even have access to their customer data, let alone the people or tools necessary to manipulate it in the way some forecasts have envisioned.

The data is the boring bit, though, so it doesn’t stop people from lusting after the promised advances, or wondering how the technologies could work for them.

Predictive analytics and machine learning are used for many things, and have been for some time. Meteorologists use those technologies to forecast the weather. Insurance companies use them to detect fraudulent activity and for underwriting. Email providers use them to power their spam filters.

These technologies help businesses identify better ways to do things, based on knowledge gathered from something that has happened before. The systems use math probability and data mining to predict the likelihood that something will or will not happen.

The technologies have many uses for marketing. For example, they can help predict a customer’s lifetime value or the likelihood that someone will respond to an offer. They can calculate the probability that a shopper will buy a certain product and refer a friend or abandon a cart and churn.

Thank you for reading this post, don't forget to subscribe to our AI NAVIGATOR!

 

3 Types of Machine Learning Algorithms

At a high level, machine learning algorithms can be placed in three groups: supervised, unsupervised and reinforced. Let’s take a look at each category:

  • Supervised learning involves an algorithm trained on data where the outcomes are already known. The correct answers exist in the historical data, and it is the algorithm’s job to recognize that answer in new data. For the purposes of marketing, this can include searching social media posts for mentions of a brand and analyzing message sentiment, or the more familiar tactic of creating product recommendations.
  • Unsupervised learning, as the name suggests, is an algorithm that has not been told what to predict. Instead, it automatically clusters into groups pieces of data that exhibit similar behavior, determining itself what is “typical” for a certain set of data. For example, this type of algorithm can identify shopping patterns that wouldn’t otherwise be obvious, and groups of people who follow those patterns. Or it can learn association patterns, such as “customers who do X often buy Y.”
  • Reinforced learning involves an algorithm that has been told several outcomes and needs to calculate an optimal course of action. For example, reinforced learning algorithms have been trained to play basic computer games, mastering what is a “good” or “bad” action in order to win the game. Generally, it is harder to use this type of machine learning for marketing, because it tends to rely on a trial-and-error approach. When you’re dealing with customer experiences, you don’t want to risk getting something wrong and damaging the relationship before you get it right. That said, reinforced learning could help refine what message and channel is best for a customer, which could improve personalized, one-to-one messaging and a campaign’s return on investment.

However, as I have suggested, there is a lot to do before you can get to this stage. In fact, I’d argue that 90 percent of the work that goes into machine learning happens before the modeling and the application of modeling take place. And that’s assuming you have the ability to collect all the relevant data in the first place. […]