How LinkedIn Levels up Your Skillset … with AI
Ever feel like your resume does not adequately reflect your vast knowledge and potential value? No worries: LinkedIn has your back.
The company goes through considerable lengths to burnish up the skillset you present to the world, using artificial intelligence, taxonomy development and lots of computer power to make sense of your profile page.
“We see a future where the world of work is centered on a skills-first economy,” wrote LinkedIn AI Tech Lead Ji Yan, leading a multi-authored blog post published Wednesday that explained how the company analyzes user data to better articulate their work skills.
Skills are everything for LinkedIn, the lingua franca of professionalism, and the great democratizer of opportunity.
For it is a person’s skills that “level the playing field in the labor market” wrote LinkedIn’s Yi Pan in an earlier post.
The skillset demonstrates “what a member is capable of — not where they went to school, where they grew up or where they worked.”
In more practical terms, a better model of a user skillset gives LinkedIn more actionable data to work with, in terms of helping that person get their next job or uplevel their expertise, as well for businesses to find better candidates, “especially in sectors that are aggressively looking for talent,” Pan noted.
A Resume Is Not Not Enough
The enthusiastic LinkedIn user will, of course, add a list of skills to their profile, in the dedicated skills section (some more thoroughly than others). But they also leave all sorts of details about their knowledge in other sections as well. They will upload resumes, and include additional information in the Summary and Experience sections. If a person takes a LinkedIn Learning course, they accrue those masteries as well.
“We don’t want to miss out on these skills, so extracting them from the text is essential,” Pan wrote.
Here is where AI comes in handy, extracting and mapping skills from all this unstructured data.
Models are fine-tuned to understand different formats, such as resumes or member profiles. “Where and how a skill is mentioned can provide a significant signal on how relevant the skill is and how we should interpret the mention of the skill,” the researchers assert.
Once unearthed, the terminology also must be normalized (will it be “data analytics” or “data analysis”?) and reconciled against LinkedIn’s skill taxonomies.
Not all skills are explicitly named in the text. Someone may write “experience with design of iOS application,” but not use the canonical name of this experience, “Mobile Development.”
A skill tagger can then do token-based and semantic-based skill matches to connect phrases to skill sets.
The semantic approach is based on a set of large language model (LLM) text encoders. The researchers name one, Multilingual BERT (M-BERT), which generates contextual embedding for source text and skill name.
Skills are also expanded through the use of LinkedIn’s Skills Graph, which can be queried for other related skills.
But Are You an Expert?
The system also takes a stab at estimating how proficient you are at a skill, a tougher job given the proclivity of people to overestimate their abilities in everything.
“While it is easy to incentivize members to list their skills on their LinkedIn profile, estimating their expertise in those skills is more challenging,” the researchers write.
They use a multitask learning framework with an “uncertainty weighting scheme incorporating signals from multiple contexts.”
This work also helps the company identify the important skills of a user, where it is used to surface opportunities that the users themselves might have missed. (To review your own Top 10 skills, log into the Skills & Endorsements section of the site.)
LinkedIn Does Real-Time Recommendations
Such data crunching across a billion users is impressive enough, but what makes this task even more challenging is that this information must always be updated.
“Without a robust tech stack for mapping content to the Skills Graph, it would just be a static list that would become outdated as time passed,” Yan et al write. “Instead, LinkedIn is able to constantly update and evolve the Skills Graph to stay current on the always-changing skills landscape.”
The results are used in many LinkedIn products and features — search, recommendations, feed rankings, job search and listings, recruiter searches, and so on.
The LinkedIn userbase makes on average, 200 global profile edits per second. Plus, LinkedIn wants each and every message processed in under 100 milliseconds.
“Serving a full 12-layer BERT model on a platform like LinkedIn, while maintaining latency standards, is a daunting task even for industry leaders since BERT, though powerful in NLP, has a large parameter count and is computationally demanding,” the researchers boast.
All this work, however, goes back to LinkedIn’s primary mission, to connect the world’s professionals and make everyone richer and more successful.
“When a job seeker opens a job posting on LinkedIn, a feature will show how many skills overlap between their profile and the job,” the researchers wrote. “In general, the higher the overlap, the more likely an application will be successful.”