How do computers optimize mass persuasion – for marketing, presidential campaigns, and even healthcare? And why is there actually no data that directly records influence or persuasion, considering it’s so important? And what’s the ideal technique for optimizing your dating life and for getting more people to wash their hands in public restrooms? It’s time for Dr. Data’s persuasion paradox “Groundhog Day”-inspired geeksplanation.
Persuasion modeling requires a “deep geek dive” – but it’s as important as it is fascinating.
Note: This article is based on a transcript of The Dr. Data Show episode. View the video here.
Of all the things achieved by machine learning, the capability to predict is the Holy Grail for driving decisions in business, healthcare, law enforcement, and more. But predict what, exactly? Like, for targeting marketing, normally it predicts which customers will buy. Then you send a brochure, to those people and only those people, the ones flagged as more likely to buy.
But hold on, a company only wants to spend $2 on a glossy sales brochure for consumers it’s likely to persuade. I mean, that’s the whole point of marketing: changing people’s minds. Influencing them. But if I mail everyone predicted to buy, I might be hitting many people who were going to buy anyway. They don’t need to be persuaded, so it’s a wasted expense, not to mention more paper consumed, more trees cut down. After all, some products just sell themselves; they fly off the shelf even with little-to-no marketing.
A New Prediction Goal: Whether They Will Be Influenced
Ok, so I’ve just convinced myself to change my prediction goal. Instead of using machine learning to predict whether a customer will buy, we’ll predict whether they’d be influenced to buy if they see this brochure. That’s a very different thing to predict.
Whoa, that does seem like a great idea. It’s a big change from traditional data driven marketing… And it applies to healthcare, also! If we’re applying a healthcare treatment based on whose health is predicted to improve, we’re making the exact same mistake as with marketing because some patients are going to improve even without treatment. If you take a pill and your headache stops, how do you know it wouldn’t have stopped anyway? And also what about predicting which patients may be hurt by a treatment, those it would actually make worse? It would be better not to treat them at all. So, instead of predicting "will the patient improve with this treatment" predict instead, "will the patient improve ONLY with this treatment — and not improve otherwise?" Will this treatment itself make a positive change?
The Dilemma: We Have No Examples of Influence in the Data
Uhhhhhh… but how can we tracked who was influenced? The only way to know someone was influenced would be if we knew that they would not have bought if we didn’t contact them, that our glossy brochure changed their mind. But we did contact them to find that out, so how could we know what would have happened if we didn’t?
Machine Learning for Persuasion: Uplift Modeling
The kind of machine learning that predicts persuasion is known as persuasion modeling or uplift modeling: Generating from data a model that predicts the influence of a treatment.
Instead of what traditional models predict — the future, the behavior, the outcome – like a purchase or an improvement in health – an uplift model predicts a treatment’s influence on that outcome.
For each individual, standard predictive modeling answers the question,“How likely is the positive outcome?” But uplift modeling answers,“How much more likely would the desired outcome be with this treatment?”
After all, the best way to do influence is to predict influence. The most direct way to know whom to market to is to know who is persuadable. Targeting in this way makes a marketing budget or a sales team more powerful.
In fact, Obama’s 2012 presidential campaign – which, you know, is just another kind of marketing campaign – used uplift modeling to improve the persuasive power of their $400 million TV ad budget by an estimated 18%, and also significantly improved the effectiveness of campaign volunteers by targeting exactly who’s door to knock on. This helped avoid knocking on the doors of “do-not-disturb” voters, which would actually backfire and inadvertently generate a vote for the opposing 2012 candidate, Romney.
After all, the whole point is to predict not just where there’ll be influence, but, more specifically, where there’ll be positive influence.
Under the right circumstances, uplift modeling improves marketing by a huge margin. Here are all the companies for which I’ve seen public disclosures of uplift modeling actually outperforming standard predictive modeling:
- US Bank
For Telenor – a big cell phone carrier in Europe – it increased a marketing campaign’s return-on-investment by a factor of 11!
How Uplift Modeling Works
Here’s one little example in marketing of how an uplift model can work. It often turns out that customers who’ve so far bought a medium amount, but not too much – those in the mid-range – are most positively influenced to buy more by marketing contact. The reason for this may be that many who’ve bought nothing at all are harder to get started, and that those who’ve already bought a lot may be negatively influenced – annoyed or otherwise turned off – if you market to them.
On the other hand, to be honest, most companies are still using standard machine learning, a.k.a., modeling, to predict outcome rather than influence – and in many cases, for good reason. Uplift modeling is more difficult. For training data, it requires the addition of a control set. And the technical methods are less well known, more complex, and more challenging to evaluate.
But I hope you agree it’s an interesting area with great potential. To learn more, check out my other more detailed article on uplift modeling, which ends with a list of links, including the real technical nitty gritty.
Request a consultation to speak with an experienced data analytics consultant to explore how uplift modeling can enhance return on your analytics investment.
This article is based on a transcript from The Dr. Data Show.