The marketer attempts to predict future behavior for targeted advertising and tailored products. With the rise of social networking sites, is it possible to use social network data from such sites as Yahoo, Google, Twitter, LinkedIn, and Facebook to improve targeting and user behavior prediction? This is the summary of an article by Sharad Goel and Daniel G. Goldstein. You can get the pdf of the behavioral targeting article here: Predicting Individual Behavior with Social Networks.
Traditionally, marketing segmentation is done by grouping individuals into segments based on their similar preferences. Television and magazine advertising, for example, can be quite wasteful because the ads reach both those who are interested and those who are uninterested with the products. However, technological advancements have made targeting more effective, especially with the rise of the online consumer. Several kinds of targeting have emerged, including demographic, location and behavioral targeting.
This paper focuses on data obtained from social networks: individual behavior records and connections between individuals.
Data and Methods
Data is obtained from the Yahoo communications network. An edge, is a connection between two people who have exchanged instant messages or email during a period of two months. That equals 719 million edges and 132 million people. Each user had a demographic profile (age and gender).
Social predictors are assessed by examining behaviors at the individual level in three domains: online and offline purchasing at a national department store chain, online recreational league participation that is played by millions, and responding to national advertisements of various services and products.
Predictions of retail purchases are based on a 1.3 million member department store customer
database, which includes one year’s purchase data for each individual. To predict participation in a recreational league, we analyzed the Yahoo Sports Fantasy Football competition, which has approximately four million annual registrants. Finally, we examine individual response to online advertisements, measured by clicks on 10
display ads prominently shown on the Yahoo front page.
Contacts of adopters are themselves considerably more likely than average to
adopt, consistent with previous investigations. For example, among retail consumers with four contacts who made a purchase during the first six-month observation period, 70% made a purchase themselves during the second
period, substantially higher than the overall purchase rate.
That contacts of adopters have relatively high adoption rates does not in itself imply that social
data are valuable for prediction.
If new data sources are to extend the limits of predictive accuracy, they must not merely substitute for traditional predictors, but should complement them. In three diverse domains, we nd that social data do in fact augment traditional methods of predicting the behavior of individuals.
Across several varied domains, we observe that social data contribute materially to predictive abilities when only basic demographics are known about a targeted base, and that this predictive advantage decreases as individual-level transactional data become available. Accordingly, social targeting seems particularly worthwhile in situations where a target’s social network is known, but past behavior|and possibly demographic information are absent.
To generalize beyond our empirical examples, we developed a theoretical model to think about new domains on two key dimensions: variation in the likelihood of adopting, and the tendency to have social ties to people with similar adoption propensities. Together, these dimensions help characterize when social data will be predictive. Our analytic results suggest social data are particularly valuable in domains where there is a small group of tightly knit, high-likelihood adopters. In this case, having an adopting contact is a strong signal that one will adopt.