Thursday, November 21, 2013

HotNets '13: AdReveal: Improving Transparency Into Online Targeted Advertising

Authors: Bin Liu (USC), Anmol Sheth, Udi Weinsberg, Jaideep Chandrashekar (Technicolor), Ramesh Govindan (USC).

Speaker: Anmol Sheth

This work is motivated by two observations: the ad ecosystem is becoming increasingly fine-grained with respect to per-user tracking, and advertisers are willing to pay (up to 2.6x higher) for targeted ads vs non-targeted ads. The goal of this work is to provide users with transparency to how these (targeted) ads are chosen. 

As an example, a person browsing a Web site sees a particular ad might want to know why this particular ad was shown to him/her. This might be useful, for example, to allow users to play an explicit role in controlling ads, for example, to separate home persona and work persona, or to not advertise my personal interests, like health, finance, or online dating.

There are three primary targeting mechanisms used today: (1) contextual (ignores user profile, but uses context, say Web page content), (2) behavioral (user profile explicitly used), (3) re-marketing (target users who have shown explicit -- perhaps previous -- interest in advertiser's product). This talk discusses how to identify whether an ad is likely to display the first or second category; more about the third in the paper.

The authors built a Chrome extension called AdReveal. It is based upon an ML model (logistic regression classifier) that is used to calculate probability of the ad using contextual or behavioral targeting. The results show that up to 65% of ad categories are behavioral. Most contextual ads are from pages about politics and government, whereas ads on pages about insurance/real estate/travel/tourism were higher. 

The authors are surprised by how high this result is for behavioral targeting.

The hope is that transparency will enable new ad control mechanisms, and promote more measurement studies. The authors also had a demo that shows the extension in action: it displays what targeting mechanism are likely used for each ad, and the breakdown of a user's ad categories. 

Q: can you use the framework to see if a user model is shared across different ad companies?
A: you mean like re-selling? not yet, but we're exploring it.
Q: we have different personas, say, across devices. These profiles may look different. Are there efforts to unify these profiles (presumably by marketers)?
A: Google seems to be trying to do this, say in android devices. Another approach is to correlate something like browsing patterns to be able to use this.
Q: how can you ensure that browsing history was consistent in experiments?
A: we used a fresh browser profile (deleted cookies, etc). we also ran experiments at different times of the day and with different IP addresses using PlanetLab.
Q: I like this work, but this seems like a passive tool; what can users to do with this information?
A: This may help users learn how to more carefully obfuscate their profile. We might be able to selectively delete profile, cookies, etc, so you don't lose the benefit of this tracking while eliminating some of the tracking.
Q: Google seems to offer this type of breakdown. Have you compared your tool to what Google reports?
A: not yet, but it sounds like a good idea.