In part one of this blog, we discussed the requirements for the foundation of an effective recommendations system: the ability to make decisions based on contextual similarity and also affinity between items based on other factors, such as user behavior, all driven by rich user data. Today, we’ll talk about how leveraging machine learning in recommendations can generate significant revenue lift by pitting algorithms against each other and learning to apply those that yield optimal results.
There are a lot of ways to power recommendations beyond simply relying on content similarity. You can trigger them based on what else the customer—or other customers in that segment—viewed, bought, or liked (customer X bought item Y and also viewed or liked item Z, etc.). You can take past behavior into account or limit the decision to the current session as well.
It’s not a “one size fits all” approach, and one algorithm is not likely to work for every use case. For instance, the women’s apparel market is different than the user intent behavior for car buyers, which is in turn different from people looking for financial services products. Each unique shopping session has a different user mindset associated with it, which is dependent on a complex set of characteristics ranging from the vertical of the organization, nature of the offers being presented, sense of urgency that is associated with the campaign, and other factors.
To optimize recommendations, Adobe Target applies multiple algorithms to different samples of your visitors, and identifies which algorithm performs the best by segment and other variables, and then automatically favors the best-performing algorithms. It also continues to explore the user behavior by ongoing experimentation. This is the goal of personalization: to show visitors the most relevant content that has the most potential to increase your revenue per visitor or conversion ratio.
Using machine learning with strategic testing in this manner can help you increase revenue even from visitors who haven’t previously purchased. Some users are simply waiting for the right product—by using optimized recommendations, you can uncover what they’re most interested in.
When evaluating recommendations solutions, look for one that will give you the flexibility to test algorithms against each other and let them “fight it out,” instead of just specifying rules. You’ll be glad you did; we’ve seen customers generate a 10% revenue lift based on recommendations alone.