Monday, November 28
12.10 - 1.00
CSE 4140

Speaker: Aditya Menon
Title: Predicting clickthrough rates using collaborative filtering

Abstract:

A basic problem in computational advertising is estimating the probability that an ad is clicked when displayed on a content publisher's webpage, a quantity known as the ad's clickthrough rate. Until recently, the state of the art for this problem has been to train a logistic regression model on carefully selected features for the webpage and ad. This talk looks at the problem from a collaborative filtering viewpoint: we treat the problem as aiming to predict the "affinity" between a particular (webpage, ad) pair - namely, the clickthrough rate of the ad on the webpage - based on the historically observed "affinities" for a collection of webpages and ads. We describe a model that predicts this affinity based on a latent feature representation for the webpages and ads, inspired by the success of such techniques in classic collaborative filtering problems. A key challenge is the extremely sparse nature of click data, which makes reliable estimation of latent features difficult. We show how this problem can be mitigated by leveraging information contained in explicit features for pages and ads. Experiments on real-world datasets show that the collaborative filtering inspired model outperforms existing approaches to this problem.