• 09.30.14

LinkedIn’s Data Science Secret: Your Hidden Org Chart

LinkedIn does more than help find sales leads or job candidates: It’s also a Rosetta stone for corporate hierarchies.

LinkedIn’s Data Science Secret: Your Hidden Org Chart
[Circles: Goldenarts via Shutterstock]

LinkedIn enjoys one big advantage against competitors Facebook and Twitter: It’s the social network people can use at work. By positioning it as a service for professional development, LinkedIn has embedded itself into offices worldwide. And in the process, found a holy grail of corporate data: The social hierarchy of people inside an organization… even if the people themselves don’t know it.


LinkedIn’s affable head of search quality, Daniel Tunkelang, spoke with Co.Labs earlier this year. Tunkelang is the person responsible for making sure LinkedIn’s searches connect people to the contacts they’re actually looking for. This means learning a lot about how people know each other, and how people interact with each other, in the process.

One thing LinkedIn’s users don’t always realize is that the search process works differently depending on whether you’re using a desktop or laptop computer, a smartphone, or a tablet. Tunkelang says typing is harder on mobile devices, which leads his team to see a higher incidence of shorter queries from users.

“In mobile, we really emphasize the autocompletion experience because the environment in which people use a laptop versus a phone is quite different.” People also use the search function differently on mobile devices too. Tunkelang told me that his company sees a lot of what he calls “meeting intelligence” being conducted on smartphones–LinkedIn users inside meetings encountering someone at a real-life event, taking out a phone, and looking up the person’s profile.

Because Tunkelang and LinkedIn’s other data scientists are able to see how users search on the service and how they use it in different circumstances, this means they get deep insights into how recruiters search for candidates, how sales teams evaluate potential leads, and how different departments of organizations relate to each other.

One of the most fascinating parts of his job, he says, is finding unexpected results when finding data to prove or disprove different hypotheses. Tunkelang’s team discovered the way people’s social networks related to each other, and found that changed the search experience. Specifically, the way people search for names on LinkedIn and the way people search for titles on LinkedIn have little to do with each other at all.

When LinkedIn users search for someone by name, it’s primarily for someone relatively closely connected to their social network (to be exact, one population away from them). But when searches are conducted for job titles, users are primarily contacting individuals two populations away from them in their social network–further away than searches by name. Although the discovery wasn’t counterintuitive, it wasn’t what they were looking for… and Tunkelang says the trend came “shining through the results” when they analyzed the data.

Other LinkedIn data projects require more user input to glean insights. Take for example that endorsement box that sometimes pops up asking you to vouch for someone’s skills? There’s actually a sophisticated project going on there.


These requests might seem like LinkedIn’s way of increasing engagement on the site, but it’s also part of a sophisticated mapping mechanism that lets data scientists figure out what job titles at organizations actually mean. Endorsements help LinkedIn figure out what skill sets and talent requirements align to which jobs.

LinkedIn engineer Sam Shah and data scientist Pete Skomoroch explained how the endorsements feature worked at the 2013 edition of data science conference Strata. Endorsements are used to build a “Skills Dictionary” for the social networking site. Defined as a taxonomy of work skills, the skills dictionary is primarily based on mining data from the site’s millions of profiles and then augmenting them through other sources like endorsements. A big part of Skomoroch and Shah’s work is cleaning the data–over 250 different phrases map to “Microsoft Office” alone.

In one case study of mapping skills to the correct occupation, they showed which phrases map to “Angels” (as in alternative medicine) and “Angels” (as in venture capital). It’s easy to figure out where psychic readings, clairvoyancy, and early-stage investing map. This information is then used to infer what skills someone with a specific title actually has.

LinkedIn then uses skills endorsements to see both which particular contacts users feel have these skills… and who they choose to endorse. The results in aggregate are used to both build social maps and to understand the difference in responsibilities between jobs with identical titles at different companies.

Facebook has its own social graph, of course. But those connections have been mapped to personal histories rather than job skills. And that results in an entirely different kind of network effect than the one LinkedIn is trying to capitalize on.