Deepak Agarwal, Shaunak Chatterjee, Yang Yang, Liang Zhang
In the 24th International World Wide Web Conference (WWW 2015)
This paper considers an application of showing promotional widgets to web users on the homepage of a major professional social network site. The types of widgets include address book invitation, group join, friends’ skill endorsement and so forth. The objective is to optimize user engagement under certain business constraints. User actions on each widget may have very different downstream utilities, and quantification of such utilities can sometimes be quite difficult. Since there are multiple widgets to rank when a user visits, launching a personalized model to simply optimize user engagement such as clicks is often inappropriate. In this paper we propose a scalable constrained optimization framework to solve this problem. We consider several different types of constraints according to the business needs for this application. We show through both offline experiments and online A/B tests that our optimization framework can lead to significant improvement in user engagement while satisfying the desired set of business objectives.