It is possible to relate an individual’s behavior and his or her social contacts. There are a few startup companies already using that strategy, and calling it social targeting. Advertising would be even more effective if social signals are used alongside behavioral targeting. This is the summary of a behavioral targeting study by Kun Liu and Lei Tang. You an get the pdf of the behavioral targeting article here: Large-Scale Behavioral Targeting with a Social Twist.
But before we can use this, there are several questions that need to be asked, and the study strives to answer these questions: 1. “Whether and how can we leverage one’s friend activities for behavioral targeting?” and 2. “Are forecasts derived from such social information are more accurate than the standard behavioral targeting models?”
Behavioral and Social Data
Both behavioral data and social data are obtained from a large information technology company in a period of two months. Behavioral data contains web browsing behavior of individuals, and categorized at the behavioral targeting level into taxonomies such as “retail.” Social data consists of nodes and edges consisting of pairs of users that authenticate themselves as buddies in an instant messaging network by the large company. Combining the two kinds of data results in 180 million users and their recorded online behaviors and those of their friends.
Homophily is a term from sociology which describes the similarities of friends in different dimensions. For example, a Microsoft research found out that friends that chatted using instant messages regularly tend to have similar characteristics and interests. Homophily is getting renewed interest because of the rise in popularity of social networking sites, especially Facebook. In fact, Facebook can identify some information regarding a user through his or her friends, even if he or she did not declare those information. These information are limited to what are called “static profiles” such as education, gender and age.
Homophily can allow us to predict the behavior of an individual based on information obtained from his or her friends. The study therefore asks the questions: “can we observe the presence of homophily in our social data, and in particular, along certain dimensions related to behavioral targeting?”
Users are qualified for behavioral targeting if they reach a certain threshold score for a given bt category. Using homophily, it is also possible to determine that a user has a bt qualification if his or her friends are qualified as well. In addition, the homophily of ad clicks is also studied, since many online publishers adopt pay-per-click.
The input for the baseline model are browsing behaviors of users, and a classifier is built which predicts a user’s chances of clicking on an advertisement belonging to a certain behavioral targeting category. Training the models is achieved using a Hadop MapReduce platform.
Leveraging Social Data for Behavioral Targeting
Supervised and unsupervised methods are developed in this study to incorporate social signals and behavioral targeting. Supervised methods are essentially behavioral targeting methods with added social features. These added features are community features and neighborhood features. Another way of combining social data and behavioral targeting is by building an ensemble classifier which combines outputs from both social and behavioral models.
The unsupervised approach uses network propagation methods to determine behavioral targeting scores directly from friends. Three types of schemes are used to do these propagation methods.
These methods were then assessed in terms of effectiveness by using large experiments with the 180 million-user network across 60 domains. This is probably the largest and most all-inclusive study regarding the benefits of social data for advertising.
The following conclusions are obtained from the experiments done in this study. First, social data can be used to compliment behavioral targeting models. Homophily of categories plays a central role in determining whether social data is effective or not, and the amount of behavioral data that a user has is directly proportional to the degree of improvement of targeting. Second, the most effective method among all the methods explored in this study is the one where social features are adjoined to behavioral targeting features directly.
The study is limited to using social data for advertising as an added source of information. There are other ways to use data obtained from social networks, like in improving the engagement that users have towards products.