News portals that simultaneously personalize every part of the landing page for every visitor and mobile health apps that adaptively tweak every part of an exercise regimen to maximize the benefit of every user are becoming plausible due to an advance in a type of interactive machine learning that my team will describe at the Annual Conference on Neural Information Processing Systems running December 4-9 in Long Beach, California. Our research falls in an important area of interactive machine learning known as contextual bandits. The area encompasses applications such as news recommendation, advertising, mobile health and digital personal assistants. These applications are implemented by computer systems that repeatedly present content to the user, detect whether the presented content is successful and over time learn to pick the best content. The fact that algorithms can optimize the content in a context-dependent manner, based on user characteristics, gives rise to the name “contextual bandits.”

At NIPS 2017 , Adith Swaminathan from Microsoft Research AI will give a talk that outlines our team’s solution to an unresolved challenge in the field: How to learn when the number of content choices, called actions, is very large – say larger than 100. We make a modest structural assumption that allows us to attack this challenge and substantially enlarge the set of applications of contextual bandits. For instance, we can now optimize not only the top article on a news portal, but the whole page. In mobile health, we can dramatically enlarge the set of possible interventions and thus improve their efficacy. Read more here…

thumbnail courtesy of microsoft.com