Facebook’s Algorithm Tweaks Show Limits of Data Science
As noted in my recent post on Publishing Executive, Facebook has been evolving its algorithm for deciding which content gets inserted into a user’s news feed. In the latest tweak, priority goes to content that is shared from the user’s personal network, and to content that “informs and entertains” and that is “authentic” (i.e. not phony clickbait). These changes have provoked widespread anxiety among content creators, given the gigantic scale of Facebook’s distribution platform. What has received less attention however is the process by which Facebook studied its users and thereby informed its algorithmic adjustments. Therein lies an important lesson for publishers as they seek to understand their own audiences.
It goes without saying that Facebook has some of the richest customer data in existence. Facebook requires individual registration and verification, so it knows me regardless of which device, app, or browser I am using -- a considerable advantage over the cookie-dependent approach that prevails on much of the web. It knows what I’m interested in, who I know, and whose posts I follow. It knows which things I find funny, which things I find alarming, which things move me to update my status, and so forth. Like the other titans of web intelligence -- Google and Amazon -- Facebook is so data rich that it is the envy of marketers and publishers everywhere. As everybody else tries to figure out how to harness and monetize their own first-party data, they realize that they still will probably be outgunned by these titans with their deep data and their deep expertise in data science.
At the same time, many of these marketers and publishers are cutting back on the conventional research functions that supported their business decisions. Why conduct a survey when you can just see what people are clicking on your website? Why run a brand tracker when you can just monitor your social media mentions? Why interview customers or conduct focus groups when you can just pay attention to what people share and like on social media platforms? Why indeed?
So it is striking that Facebook, the company at the apogee of Big Data, made extensive use of traditional research methods in studying users and informing the algorithm changes. For example, earlier this year it recruited samples of users and then showed them a series of A/B alternatives, asking them which they would most like to see in their news feed. They then compared their choices to what the algorithm would have given them. They even (gasp!) asked the participants direct questions probing the reasons for their preference -- a very old-school approach!
The exercise showed that users often valued content that was serious or sad, but that this was not content that they would routinely share or like or comment on. Since the algorithm was so tuned to content that generated a direct response, it was delivering a sub-optimal experience to users. The discovery led Facebook to consider algorithm adjustments to better sniff out content that truly informs and entertains, or that is regarded as “authentic” -- even if it does not generate a lot of observable sharing/liking/commenting action.
Similarly, Facebook used a traditional research approach -- a five country survey of 2,000 users, conducted by Ipsos Mori -- to understand the attitudes of those who use commercial ad blockers before making this week’s dramatic announcement that Facebook’s desktop version would henceforth defeat ad blockers.
I am not suggesting that Facebook is placing any less priority on data science or that it will no longer take advantage of its gigantic observational window on how its users behave. But in an era when marketers and publishers are busy chasing the shiny new toy of Big Data, it is striking that a company like Facebook recognizes the value of combining both the data sciences and the venerable workhorse tools of market research. Wise publishers would do the same.
Related story: Publishers Have Only Scratched the Surface of Audience Data