TL;DR
This paper introduces skills2graph, a novel system that leverages a vast corpus of online job vacancies (OJVs) and a knowledge-rich taxonomy to address the persistent job skill mismatch. By combining distributional semantics from word embeddings with count-based skill relevance measures, skills2graph provides personalized job recommendations and insights into skill gaps for individuals and labor market analysts.
What Problem Does This Paper Address?
The modern job market is characterized by an overwhelming number of job postings and applicants, leading to a significant "skill mismatch" problem. Traditional approaches to job recommendation and skill assessment often rely on handcrafted features and expert knowledge, making them costly, difficult to update, and prone to errors. This paper tackles the challenge of efficiently and accurately matching individuals' skills to suitable job opportunities in a dynamic labor market, especially across different countries and industries. They aim to provide a data-driven, adaptable solution that can help individuals understand their current standing, identify reskilling opportunities, and plan career paths.
Methods
skills2graph employs a multi-faceted approach:
- Taxonomy-Aware Word Embeddings: They train word embeddings (using FastText) and evaluate them intrinsically using a novel automated measure called HSS (Human-crafted Similarity Score). This method aims to capture semantic similarity between taxonomic elements (like skills and occupations) more effectively than previous approaches, which struggled with word sense disambiguation and the hierarchical nature of taxonomies like ESCO.
- Skill-Occupation Relevance (RCA): They compute a count-based measure called Revealed Comparative Advantage (RCA) to quantify the importance of a skill for a given occupation. This is adapted from previous work but uses skill frequency from OJVs since the ESCO taxonomy doesn't provide a direct measure. They also normalize RCA to a [0,1] range for easier interpretation.
- Graph Database (S2JGraph): The extracted information (embeddings and RCA) is stored in a directed labeled multi-graph database called S2JGraph. This allows for efficient querying and traversal to answer complex questions about skill-job relationships.
- Occupation Ranking: Given a user's skill set, occupations are ranked based on a combination of cosine similarity between the user's skills and the occupation's required skills, and the RCA of those skills for the occupation.
Datasets
- Online Job Vacancies (OJVs): A large corpus of 2.5 million+ OJVs from the United Kingdom, France, and Germany.
- European Standard Taxonomy of Occupations and Skills (ESCO): A knowledge-based taxonomy used to structure and understand the relationships between skills and occupations.
Key Insights and Ideas
- Bridging Distributional and Knowledge-Based Approaches: The paper effectively combines the power of distributional semantics (from word embeddings) with structured knowledge (from ESCO and RCA) to create a richer representation of skill-job relationships. This hybrid approach is more robust than relying on either method alone.
- Automated Semantic Similarity for Taxonomy Evaluation: The development and use of the HSS measure for evaluating word embeddings against taxonomic elements is a significant contribution. It offers a more scalable and adaptable alternative to human-annotated datasets for this specific task.
- Data-Driven Skill Relevance: By using RCA derived from actual job postings,
skills2graphcaptures the real-world demand for skills, making its recommendations more practical and up-to-date than systems relying solely on static expert knowledge. - Graph Representation for Complex Queries: Storing the data in a graph database is crucial for enabling sophisticated queries that can answer nuanced questions about career paths, reskilling, and upskilling.
Key Results
The skills2graph system was evaluated through a user study with 10 labor market experts. The results were highly positive:
- High Precision@3 (P@3): Achieved 0.823 for P@3-3 and 0.610 for P@3-4, indicating that the system's top recommendations were frequently considered relevant by experts.
- Excellent Normalized Discounted Cumulative Gain (nDCG): Achieved an nDCG score of 0.985, demonstrating a strong correlation between the experts' scores and the system's ranking of recommendations. This suggests the system effectively prioritizes the most relevant job suggestions.
Limitations and Future Work
- Country-Specific Models: While the system is designed to be adaptable to different countries, the current implementation and evaluation are based on data from the UK, France, and Germany. Further validation across a wider range of countries and languages would be beneficial.
- User Study Scope: The user study involved a small group of labor market experts. Expanding the user study to include a broader range of users (e.g., job seekers, career counselors) could provide more diverse feedback.
- Dynamic Labor Market: The labor market is constantly evolving. While the system can be updated, continuous monitoring and adaptation to new skills and job trends will be essential for long-term effectiveness.
- Future Work: The authors suggest extending the system to support more complex queries and to integrate with other labor market intelligence tools. They also highlight the potential for using
skills2graphto analyze broader labor market trends and policy implications.
Summary updated