This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Audience Selection for OnAudience Selection for On--line Brand Advertising:line Brand Advertising:PrivacyPrivacy--friendly Social Network Targetingfriendly Social Network Targeting
Foster ProvostFoster Provost
withwith Brian Brian DalessandroDalessandro, Rod Hook, Rod Hook, , XiaohanXiaohan ZhangZhang, Alan Murray, Alan MurrayThis work conducted while the authors were at Media6°, Inc.
• What information can firms know about you based on your on-line browsing behavior?
• Who you are? • Your demographics?• Psychographics?• Income? • Your medical conditions?• Where you live and your kids’ names? • Every page that you visit?
Audience Selection for OnAudience Selection for On--line Brand Advertising:line Brand Advertising:PrivacyPrivacy--friendly Social Network Targetingfriendly Social Network Targeting
Foster ProvostFoster Provost
withwith Brian Brian DalessandroDalessandro, Rod Hook, Rod Hook, , XiaohanXiaohan ZhangZhang, Alan Murray, Alan MurrayThis work conducted while the authors were at Media6°, Inc.
• Non-premium display ad market predicted to grow significantly faster than the rest of on-line advertising (e.g., sponsored search, premium display, contextual)
– (Coolbrith 2007) – largely due to the stabilization of the technical ad-serving
infrastructure based on the consolidation into a small number of ad networks (e.g., Doubleclick, RightMedia)
• There is evidence that display brand advertising increases purchases (on-line and off-), and improves search advertising as well (Comscore 2008, Atlas Institute 2007, Fayyad personal communication, Klaassen 2009)
Hill, Provost, and Volinsky. “Network-based Marketing: Identifying likely adopters via consumer networks. ” Statistical Science 21 (2) 256–276, 2006.
Prior work:
Social network targeting (Hill et al. ‘06)
• Define Social Network Targeting--> cross between viral marketing and traditional
• target “network neighbors” of existing customers• based on direct communication between consumers• this could expand “virally” through the network without any word-of-
mouth advocacy, or could take advantage of it.
• Example application:– Product: new communications service– Firm with long experience with targeted marketing– Sophisticated segmentation models based on data, experience, and
intuition• e.g., demographic, geographic, loyalty data• e.g., intuition regarding the types of customers known or thought to have
affinity for this type of service
• Results: tremendous lift in response rate (2-5x)
1. A privacy-friendly technique for selecting and targeting brand audiences:
– by finding social neighbors of existing “brand actors”– using no-PII at all– based on visits to social-networking pages (and other UGC)– ironically, a 3rd-party ad network can provide greater privacy
2. A method for evaluating on-line brand audiences– based on density of “brand actors”– following the ideas from hold-out evaluation for predictive
modeling
3. A demonstration that the network neighbor audiences indeed have strong brand affinity
UGC = user-generated content (non-professional)brand actors = browsers having taken an action associated with brand affinity
• Define two time periods: t1 and t2– all decisions on selecting audience A are made during t1
– t2 is disjoint from and subsequent to t1
• Collect brand action takers in t2 (disjoint from seeds)
– call this set: B2+
• To evaluate an audience A we can compute the future density of brand actors:
• We would like to know how well our brand proximity measures rank future brand actors, so we can compute the area under the ROC curve (AUC) for any measure
(from a working ad network)• a sample of about 10 million anonymized browsers • all of their observed visits to social networking content over 90 days (from several of the largest SN sites)
• bipartite graph: – 107 x 108 with ~2.5 x 108 non-zero entries
• quasi-social network: – 107 nodes with 20-40 neighbors each (on average)
• Resultant audiences per brand– on average ~100K seed nodes – total network neighbor audience pool: 2-4 million
We selected a small set of high-ranking network neighbors for three group-2 brands. In production we showed them only public service announcements (PSAs). We did the same (with the same campaign parameters) for a “run of network” campaign (bid on everyone).
We acquired from the ad exchange the rates of conversion -- here “organic” conversion.
• We can build high brand-affinity audiences by selecting the (quasi-) social network neighbors of existing brand actors, identified via co-visitation of social-networking content.
• These neighbors take brand actions at a higher rate organically, as well as after being targeted by ads, and the highly ranked neighbors do especially well.
• We can learn better models by combining evidence from the individual brand proximity measures.
• The quasi-social network likely embeds a social network.
Primary concern: people (say employees) would have access to sensitive, harmful, or embarrassing information about you.
Secondary (but nonetheless important) concern: due to data breaches these data would become public.
The privacy-friendly technique:• has no need to ever keep PII• does not keep information about content either• does not use user-supplied SN profiles (in contrast to..)
--> thus, addresses both concerns
• plus, seems to be relatively safe from reidentification techniques