Did you miss a session from the Future of Work Summit? Head over to our Future of Work Summit on-demand library to stream.
LinkedIn today pulled back the curtains on Qualified Applicant (QA), an AI system that learns from job candidate interactions the kinds of skills and experience a hirer prefers. It’s the model the Microsoft-owned platform uses to help over 690 million users in 200 countries find jobs for which they have the best chances of hearing back, and which aims to reduce the likelihood recruiters overlook applicants by highlighting those deemed a fit.
Creating a system that can contend with the transient nature of job posts was no walk in the park, according to LinkedIn. It had to work at scale — QA has “billions” of coefficients — and it had to be effective for as many job seekers and hirers as possible. Formally, QA tries to project the probability of a “positive recruiter action” conditional on a given member applying for a specific role. What constitutes a positive recruiter action depends on the context — it can include viewing an applicant’s profile, messaging them, inviting them to an interview, or sending them a job offer.
The single global QA model is individually tailored to members and roles, with per-member and per-job models trained on data unique to the members and jobs. Each of the many models is independent within a single training iteration, making them parallel and easier to serve at scale. While the global model is trained on all data, per-member models are trained using only members’ job applications. Per-job models, meanwhile, are trained on jobs’ applicants.
The global QA is retrained once every few weeks, but the personalized models must be refreshed regularly to combat degradation. (LinkedIn says the per-member models’ performance advantage over the baseline halves after three weeks.) Training labels are generated every day from events like hirer engagement with new candidates; an approximate label collection pipeline heuristically infers negatives and uses explicit positive and negative feedback as soon as it becomes available. For example, if a recruiter responds to other applications submitted later, the pipeline might infer a negative label for an application with no engagement after 14 days.
It takes up to a day to generate labels and retrain the personalized QA model components, which are only deployed if they pass certain automated quality checks. In the future, LinkedIn hopes to reduce the lag time to minutes with a near-real-time data collection and training framework built atop stream processing technologies like Apache Samza and Apache Kafka.
Across LinkedIn business lines where QA has been deployed — Job Seekers, Premium, and Recruiter — the company says it’s enabled new experiences. On the seeker side, QA highlights search results if a member’s profile is a good match for the job. For Premium members, it showcases opportunities for which members are competitive with other job applicants. And hirers using LinkedIn Recruiter benefit from a smarter ranking of applicants, as well as notifications for members with very high match scores.
LinkedIn says the personalized models delivered “double-digit” gains in hirer interaction rates and click-through rate (CTR) for recruiter notifications compared with the systems they replaced, as well as a “site-wide lift” in confirmed hires and premium job seeker CTR. “Our analysis demonstrates that the majority of job applicants apply to at least 5 jobs, while the majority of job postings receive at least 10 applicants. This proves to result in enough data to train … personalization models,” LinkedIn wrote in a blog post. “Our vision … is to create economic opportunity for every member of the global workforce. Key to achieving this is making the marketplace between job seekers and hirers more efficient … Active job seekers apply for many jobs, and hear back from only a few.”
LinkedIn’s use of AI is pervasive. In October 2019, the Microsoft-owned platform revealed a model that generates text descriptions for images uploaded to LinkedIn, achieved using Microsoft’s Cognitive Services platform and a unique LinkedIn-derived data set. LinkedIn’s Recommended Candidates feature learns the hiring criteria for a given role and automatically surfaces relevant candidates in a dedicated tab, and its AI-driven search engine employs data like the kinds of things people post on their profiles and the searches that candidates perform to produce predictions for best-fit jobs and job seekers. Moreover, LinkedIn’s AI-driven moderation tool automatically spots and removes inappropriate user accounts.
VentureBeat's mission is to be a digital town square for technical decision-makers to gain knowledge about transformative enterprise technology and transact. Learn More