Serving Up Automated Personalization: Revealing the Secret Recipe

In my last post, I talked about how Target’s Automated Personalization (AP) has changed the rules of digital marketing. My colleague, Jonas Dahl, also published a great article on its benefits. Although rules-based targeting still makes sense in certain scenarios with very defined audiences, Automated Personalization takes targeting to a new, dynamic level that simply can’t be achieved manually.

But how does it work? How is it possible for a single tool to tackle such complex logic all on its own?

Revealing the “Secret Recipe”

With other personalization solutions, marketers aren’t privy to what’s happening behind the scenes. They have no choice but to blindly trust the criteria and calculations, hoping that the right targeting decisions are being made. But at Adobe, there is no “black box.” We want our customers to know exactly how their tools work, so they can have full confidence in the results.

Step 1: Gather Your Ingredients

If you think of AP as a recipe, there are several core ingredients that go into it:

Together, all of these ingredients constitute thousands of data points, which are put into one of the Automated Personalization “blenders” and mixed together to create the perfect content recipe.

Step 2: Choose Your Blender

With Automated Personalization, marketers can choose from two different types of “blenders”: the random forest model and the residual variance model (RVM). Each has its own distinguishing features and advantages, and each produces different content results.

The Random Forest Model

Random forest is a custom model built on the state-of-the-art Hadoop/Mahout platform. It combines data about the individual visitor with historical data from similar visitors to calculate the probability that a consumer will take the desired action—whether that’s clicking, buying, or subscribing.

After building hundreds of decision trees based on random data splits, the model calculates an average prediction across all of the trees. This determines the most predictive variables at the offer level and which content will be served up to the user in real time.

Random forest offers an extremely high rate of accuracy: up to 46 percent higher than machine-learning applications that use the logistic regression method. Its speed of learning is another key benefit, as the model performs quick calculations to accommodate fast-changing campaigns. Random forest can effortlessly handle even the largest data sets, seamlessly incorporating thousands of variables.

Residual Variance Model

The legacy RVM offers a simpler infrastructure, tracking a total of twenty variables with a single decision-tree model.

RVM’s simpler structure is best suited to long-lasting campaigns that don’t have much deviation.

Step 3: Get the Perfect Blend of Personal & General

Whatever type of personalization recipe you’re using, the end result depends on a combination of personal and general data.

The personal ingredients include things like previous purchases, location, device, past behaviors, and other individual characteristics that determine the likelihood of converting. To further improve the predictive power of Automated Personalization, Adobe Target uses advanced Web-crawler technology to segment the entire website into interest areas. This allows the algorithms to segment people according to which pages or categories they visited during prior site visits. Generalized ingredients include the collective behavior of all visitors, to determine how the crowd responds to a particular offer.

Every time a visitor comes to a site, AP looks at both general and personal data. Using a “multi-armed bandit” algorithm, the model performs a split-second decision of which content is most likely to lead the visitor toward the desired outcome. This is especially effective when a campaign is first launched or when a new experience is added to an existing campaign. In a nutshell, the multi-armed bandit provides the optimal balance between learning from past experiences and exploring the potential for new experiences.

Step 4: Test the Recipe As You Go

Whereas traditional personalization methods are tested after the fact, Automated Personalization incorporates testing as part of its approach. This allows marketers to measure the effectiveness of an experience in real time, and then make the appropriate adjustments to boost revenue immediately—not days or weeks later.

Now that you know the basics about how Automated Personalization works, you may be wondering how all of this information reaches you—and how to incorporate all of the insights into your day-to-day marketing practices. In my next post, we’ll peel back another layer of Automated Personalization to explore its robust reporting capabilities.