On the back of my business card it says, “Baynote: personalized shopping experiences.” This can often seem very remote from what we in engineering work on day-to-day, where much time is spent in the technical guts of the products. So every now and again it’s good to think about the end goal that we’re working towards – providing services to our customers that allow them to personalize their websites to each visitor, and deliver a compelling customer experience. One aspect of this is recommendations, and while there may be value in simply displaying product recommendations, in that it makes the site more engaging or easier to navigate, our customers are also interested in their bottom line. How do we determine which recommendations will have the most impact for a customer?
All recommendations are not created equal. Some recommendations are clicked on more than others – they have a higher click-through rate. Some items are purchased more often once viewed than others. They have a higher “buy-through rate”. And some items have more value to a customer once purchased than others. The simplest measure of value is revenue; some items cost more than others. Yet, from a business perspective, margin may be more important than revenue. And in certain cases a customer may be more interested in other key performance indices (KPI) – they may be interested in acquiring new customers, or they may have a warehouse full of a seasonal item that they would like to sell before the end of the season.
In my last post, I discussed how Bayes’ Theorem allows us to learn from data. It does this by modeling knowledge using probability distributions. We learn a probability distribution over the potential recommendations from the data we collect. We can use this probability distribution directly to determine which recommendations to show. But when a customer also tells us which KPI they are most interested in optimizing, we can combine the KPI information for each item with this probability distribution. We then rank recommendations not by the probability that a user will purchase the recommended item, but instead by the expected value of the KPI. This results in recommendations that will have the most impact on the metric that each customer has told us is most important to them.