סמינר מערכות מידע

הסמינר יינתן ע"י:  מר יונתן גור, מאוניברסיטת קולומביה, ארה"ב

Abstract: 

A new class of online services allows publishers to direct readers from online articles they currently read to other web-based content they may be interested in. We study the dynamic content recommendation problem, focusing on the questions of how to maximize the number of clicks along the path of a reader. Based on a rich data-set, we develop a representation of content along two key dimensions: clickability, the likelihood to click to an article; and engageability, the likelihood to click from an article. We propose a class of look-ahead heuristics and show, through a simulation and theoretical bounds, that engageability is a key to capture significant part of the performance gap between optimal recommendations and myopic ones that are used in current practice. We propose an approach to implement these heuristics “on the fly” based on click history, without increasing the complexity of the current process. The impact of using the proposed class of recommendations is being tested in a controlled experiment designed with our collaborator, a leading provider of dynamic content recommendations. When only a little history is available, we suggest an approach to recommend articles by tracking attributes of their main topic. Our approach is tested and shown to outperform current methods for recommending new articles.

Joint work with Omar Besbes and Assaf Zeevi, Columbia Universit