Every day, online ad agencies need to answer a crucial question: Which ad is more likely to compel viewers to click? And they must answer this question billions of times, for millions of users on thousands of news websites, streaming websites, social media platforms, and mobile apps.
With billions of ads being served every day, being able to improve click-through rates by a fraction of a percent can have a significant impact on revenue. How do they do it?
In my latest “Deconstructing AI applications” article, I explore the role of reinforcement learning in ad optimization. First, we discuss A/B/n testing, the classic method for comparing different solutions. Then we will learn how “multi-armed bandit” algorithms, a type of single-step reinforcement learning, can provide an efficient and dynamic alternative to A/B/n testing. Finally, we’ll look at ways to further improve MAB models and take a peek at “contextual bandits,” a type of RL that can personalize ads for every user.
Read the full article here.
If you’re interested in exploring more AI applications, take a look at the other articles in the “Deconstructing AI applications” series.