How our AI is helping members navigate the Great Reshuffle

October 27, 2021

Co-authors: Liangjie Hong, Alexandre Patry, Benjamin Le, and Fei Chen

Spurred on by the COVID-19 pandemic and the many changes it has brought, we’re currently in the midst of what we’re calling the Great Reshuffle: a period of transitions in which people are rethinking why and how they work. Nearly 90% of professionals say that the pandemic caused them to reevaluate their priorities in work and life, and as a result, many are looking for new opportunities. According to our global data, job transitions are up 50% year over year. As more candidates are looking for new positions, we’re seeing a record-setting number of open jobs on LinkedIn. Remote jobs are also becoming more prevalent, with 1 in 8 U.S. jobs on LinkedIn now being remote, an increase over the previous ratio of 1 in 67 seen in March 2020.

With so many potential job opportunities for candidates, and such a large talent pool for hirers, finding the right match might seem more difficult than ever—especially with the rise of remote work expanding the options for both groups. However, our data shows that when job seekers explore open roles on LinkedIn, they are finding jobs roughly 6% faster than before, and with fewer applications. Ultimately, this dynamic can signal a more efficient matching process between talent and opportunity in a tight labor market. One way that we aim to make finding the right fit easier for both hirers and candidates at LinkedIn is by augmenting recruiting and job searching workflows with artificial intelligence (AI).

We’ve long used AI in our LinkedIn Talent Solutions (LTS) platform, but the unique scale of the Great Reshuffle makes it even clearer how helpful AI can be in matching people, skills, and job openings. And we’re excited that, thanks to several innovations over the past few quarters, our AI for LTS is performing better than ever. In the past five quarters, we’ve improved predicted confirmed hires (PCH) by 34.4% in LTS due to our AI efforts.

We define PCH as the likelihood of an action (like applying to a job or sending an InMail) on LinkedIn leading to a member being hired. To confirm the accuracy of the PCH model, our systems periodically analyze matured confirmed hire data and compare it to PCH estimates. In addition to the increased PCH for the LTS platform as a whole, we’ve also seen PCH improvements in Jobs You May Be Interested In (about +14%) and Job Search (about +11%) thanks to our AI updates.

Below, we’ll cover some of the primary AI innovations that have led to these improvements in the efficacy and accuracy of our Talent Solutions platform.

Pensieve: An embedding feature platform

An important part of building AI models is feature engineering, which consists of introducing new features that make raw data easier for AI models to read. Pensieve is an embedding feature platform developed for the Talent Solutions and Careers product to pre-compute and publish entity embeddings. Embeddings are useful in feature engineering because they compress a large amount of high dimensional information into a dense representation, which makes them easier to leverage as features in models. Pensieve embeddings are trained with supervised deep learning techniques using semi-structured data from members’ profiles, job postings, and other sources, and they are used by ranking models in latency-sensitive applications across Talent Solutions and Careers. By pre-computing the embeddings, we avoid the latency challenges typically seen with the compute-heavy forward propagation of representation learning. We’re continuing to develop the Pensieve platform, including by incorporating raw text features into the model during embedding training. We’ve also found ways to use Pensieve to represent member job seeking activity by building models that learn to take the list of Pensieve embeddings of jobs that a member applies to and summarize that list into a single member job seeking activity embedding. 

In the last year, ramping Pensieve has led to approximately 10% PCH improvements combined across various LTS products.  

Standardization of data representations

Standardization refers to the process of standardizing ambiguous or heterogenous raw data, like members’ titles (e.g., “Software Engineer,” “SWE,” and “Sr. SW Eng”), to semantically well-understood representations in the LinkedIn Economic Graph. It’s achieved through a combination of human curation and advanced AI approaches like using graph neural networks and deep transfer learning. Standardization lays a solid data foundation for all of our LTS products and plays a vital role in matching members to relevant opportunities. Examples include: 

  • Standardizing members’ inputs, like job titles in Job Search, through the standardization typeahead services 

  • Standardizing key entities like Skills in both member profiles and job postings to enable meaningful and explainable matches between the two

  • Powering reliable and actionable insights for members, job seekers, recruiters, and labor market decision makers

In the last year, by improving data quality, the Standardization team contributed to an approximately 8% lift in PCH across LTS.

Continuing our AI work for LinkedIn Talent Solutions

More than a billion data points make up the LinkedIn Economic Graph, which consists of all of the jobs, skills, people, companies, and schools—and the relationships among them—represented on our platform. We believe that every member and customer can benefit from insights derived from this data, and AI is what allows us to operationalize those insights at scale. With AI, we can create personalized job recommendations for each member, or surface the most relevant candidates for an open position to recruiters.

As we incorporate AI into our LTS platform, we’re also committed to ensuring that we’re practicing our responsible AI values. To that end, we always look for ways to improve equity in our Talent platform, such as our efforts to build representative talent searches into LinkedIn Recruiter. 

We’re proud of the work our teams have done to further improve the AI that powers our jobs and recruiting platforms, and we’re excited to continue innovating in this space to create the best possible experience for our members and customers. During the Great Reshuffle and beyond, we believe that AI can augment the efforts of job seekers and hirers to help match them to the most relevant opportunities amid a deep sea of options.