"The Fault in Our Recommendations: On the Perils of Optimizing the Measurable" The main idea is simple, and yet, we think, powerful. I'm going to give a highly informal version of it here. Online platforms for content (or goods and services) typically optimize recommendations for engagement (or purchases). This is natural since engagement is easily and immediately measurable, and the longer you can extend each session of the user, the more ads and sponsored content you can show them, right? Not quite, we argue. All clicks are not created equal. Some generate a lot more value for the user than others, and facilitating such discovery could boost user retention and lifetime value. Discovering "niche" content that they like can generate a lot more value for some users than only consuming mass market content, while other users prefer sticking to mass market content. Crucially, modern platforms have the ability to simultaneously serve both kinds of users well, even without knowing a priori which users are interested in exploring niche content, and even without being able to precisely measure user utility. The way to do this, we argue, is to recommend users a mix of niche and mass market content. Such an approach minimally reduces engagement while boosting user welfare significantly, relative to an engagement maximization approach which only recommends mass market content.
I should mention that Akshit is a rising final year PhD student in Columbia DRO, co-advised by Omar and me, and he'll be on the 2024-25 faculty job market. He's great, keep an eye out for him! Ping me if you want to know more about him.
P. S. There is some energy now around utility-aware recommendations. Here's an interesting workshop on being organized by my colleague Dan Russo and others at Recsys 2024: https://sites.google.com/view/sure-2024/home
