SMITH BRAIN TRUST — Company pages on social media sites like Facebook contain real-time information for brand managers hoping to gauge customer sentiment, but only if they know how to make sense of the raw data. New research from the University of Maryland’s Robert H. Smith School of Business shows how to filter for bias, starting with a three-pronged test to identify and remove fake users.
Bloomberg estimated in 2012 that up to 40 percent of all social media accounts are fake. A more recent report from the University of Iowa and other institutions shows how fake Facebook fans are sold by the thousands.
“For comments on Facebook brand pages to reflect genuine user experiences, opinions and interactions with brands, such fraudulent activities should be detected and removed,” write Smith School professors Kunpeng Zhang and Wendy W. Moe, authors of the working paper.
Staying ahead of counterfeiters is not easy, which is why the U.S. Treasury continues to add safeguards to paper currency. In the social media realm, not even Facebook knows how many of its users are fake. Its 2013 annual report estimates a range of 70 million to 140 million fakers — a population somewhere between the size of Thailand and Russia. That's a wide range, like a weather forecast with a high somewhere between 40 and 80 degrees Fahrenheit.
Zhang and Moe, director of the Smith School’s Master of Science in Marketing Analytics, limit the guesswork by examining more than 7 billion Facebook “likes” and comments from 170 million users on more than 3,000 company pages. Using this massive database, the authors define normal user activity and then flag outliers in three categories.
Suspicious quantity: Typical users in the study commented on four to five pages and “liked” posts on seven to eight pages. But one suspicious user appeared on more than 600 brand pages, and another user “liked” posts on more than 500 brand pages. For the purposes of the study, the authors ignored users who engaged on an extremely large amount of brand pages, such as commenting on 100 or more brand pages or “liking” posts on 150 or more brand pages.
Suspicious ratio: Typical users “like” less than 1 percent of posts on a brand page. But Moe and Zhang found one suspicious fan who liked 7,963 of 8,549 posts for the same brand — more than 93 percent of the total. That might be a proud parent of the CEO, but unlikely. Users in the study who “liked” more than 90 percent of posts from the same brand were ignored. “Most loyal users are still under this threshold,” the authors write.
Suspicious duplication: Moe and Zhang also filtered users who posted duplicate comments containing url links, which sometimes direct to phishing sites. One test on CNN’s page found 237,101 duplicate comments out of 12.5 million — about 2 percent of the total.
Zhang says the study applied simple and conservative rules to identify fake users. A more aggressive approach might work better for specific companies. The key is to identify outliers in each category.
Once a company scrubs fake users from Facebook and other social media platforms, additional steps remain to address potential bias. Moe and Zhang lay out a complete approach in their study, which produces results similar to traditional brand tracking surveys — but with less cost and delay.
“Companies can clean the data,” says Moe, co-author of an earlier book, Social Media Intelligence (Cambridge University Press, 2014). “They can move from social media monitoring to social media intelligence.”
Read more: Bias on Your Social Media Brand Page: How Do We Measure Underlying Brand Favorability and Identify Potential Bias in Sentiment Metrics? is a working paper.
GET SMITH BRAIN TRUST DELIVERED
TO YOUR INBOX EVERY WEEK