Predictive Client-side Profiles for Keyword Advertising

  • Mikhail Bilenko ,
  • Matthew Richardson

NIPS 2010 Workshop on Machine Learning in Online Advertising |

Current approaches to personalizing online advertisements rely on estimating user preferences from server-side logs of accumulated user behavior. In this paper, we consider client-side ad personalization, where user-related information is allowed to be stored only within the user’s control (e.g., in a browser cookie), enabling the user to view, edit or purge it at any time. In this setting, the ad personalization task is formulated as the problem of iteratively updating compact user profiles stored client-side to maximize expected utility gain. We compare the performance of client-side profiling to that of full-history server-side profiling in the context of keyword profiles used to trigger bid increments in search advertising. Experiments on real-world data demonstrate that predictive client-side profiles allow retaining a significant fraction of revenue gains due to personalization, while giving users full control of their data.