Abstract
At LinkedIn, connecting the global workforce with economic opportunity requires sophisticated matching systems. Two primary engines drive this: a reactive Job Search system and a proactive "Jobs You May Be Interested In" (JYMBII) recommendation system. A critical component of these systems is the extraction of relevant skill entities from job postings. Traditional text-based approaches often rely on Named Entity Recognition (NER), which is typically salience-blind and market-agnostic. This work introduces Job2Skills, a deployed system that incorporates multi-resolution salience and market dynamics—such as supply and demand—to identify the most impactful skills for job targeting. Job2Skills has demonstrated significant improvements in online job application rates (+1.92%) and a substantial reduction in skill suggestion rejection by recruiters (-37%).
The Challenge: Beyond Simple Keyword Matching
When a recruiter creates a job posting, the platform must identify the core skills required to target the right candidates. Historically, this was treated as a standard NER task: identifying every skill mentioned in the text. However, this approach suffers from two major limitations:
- Salience Blindness: Traditional NER treats all mentions equally. In a posting for an AI Researcher, "Communication Skills" might be mentioned alongside "Deep Learning." While both are present, only the latter is a salient, core requirement for targeting.
- Market Agnosticism: Standard extraction ignores the broader labor market. A skill might be mentioned in the text but be so common that it fails to differentiate candidates, or so niche that it restricts the applicant pool too severely. Effective targeting requires understanding the "supply" (how many members have the skill) and "demand" (how often it appears in similar jobs).
The Job2Skills Framework
To address these gaps, Job2Skills treats skill extraction as a utility optimization problem. The system utilizes an XGBoost classifier fed by a diverse set of features designed to capture both the textual importance and the market context of a skill.
Multi-resolution Skill Salience
The model evaluates the importance of a skill candidate across three distinct resolutions to ensure it is central to the job's function:
- Sentence-level: Using encoders like BERT or FastText, the model analyzes the specific context in which a skill is framed.
- Segment-level: The system groups mentions by section (e.g., "Requirements" vs. "Benefits") and measures how well a skill aligns with the thematic "centroid" of that segment.
- Job-level: The model compares the skill candidate against all other entities and titles mentioned throughout the entire posting to determine global relevance.
Incorporating Market-Aware Signals
A key innovation of Job2Skills is the integration of LinkedIn’s massive internal labor market data. The model calculates the "Member-side supply" (prevalence in the workforce) and "Job-side demand" (frequency in similar postings). By using metrics like Point-wise Mutual Information (PMI) and Entropy, the system identifies skills that are uniquely characteristic of specific professional cohorts, ensuring that targeting is both precise and reachable.
Datasets and Training
Because no standard dataset for "salient" skills existed, the researchers developed two novel ground-truth sources:
- Job Targeting (JT) Dataset: This utilizes explicit feedback from recruiters. When a recruiter accepts or rejects a suggested targeting skill during the posting process, they provide high-quality "expert" labels.
- Quality Applicant (QA) Dataset: This uses distant supervision by tracking members who applied for a job and received positive recruiter feedback. The skills shared by these successful applicants serve as a proxy for salient requirements.
Results and Impact
The deployment of Job2Skills has yielded measurable improvements across LinkedIn’s ecosystem:
- Offline Performance: The model achieved a +62.67% relative improvement in AUROC compared to previous market-agnostic baselines.
- Online Job Recommendations: In live A/B testing, the system drove a +1.92% increase in job applications and a +6.71% increase in member coverage, meaning more members were matched with relevant opportunities.
- Recruiter Experience: The rejection rate for skill suggestions provided to job posters dropped by 37.06%, confirming that the extracted skills better align with recruiter intent.
Conclusion
Job2Skills represents a shift from simple entity recognition to utility-based entity extraction. By combining deep learning for textual salience with real-world market signals, LinkedIn can more effectively bridge the gap between job requirements and member expertise. Future work will focus on the temporal dynamics of skills—tracking how the importance of specific technologies or competencies shifts over time—to keep the targeting engine aligned with the evolving global economy.
Summary updated