Skills2Job: A Recommender System that Encodes Job Offer Embeddings on Graph Databases (Student Abstract)

Andrea Seveso, Anna Giabelli, Lorenzo Malandri, Fabio Mercorio, Mario MezzanzanicaAAAI 2021 · DOI

Abstract

Skills2Job is a data-driven recommender system designed to map user skills to the most suitable job vacancies across different countries using word embeddings and graph databases. Beyond simple matching, the system identifies specific "skill gaps," providing users with a roadmap of the competencies required to transition into a target role. By leveraging millions of online job vacancies and hierarchical taxonomies, the system offers a scalable, automated alternative to traditional expert-defined recommender systems.

Introduction

Matching the right person to the right job is a significant challenge in the modern labor market. Traditional recommender systems often rely on "handcrafted" features and expert-defined rules. While accurate, these systems are expensive to build, difficult to update as the market evolves, and prone to human error.

The authors address the need for a "knowledge-poor" (data-driven) approach. They propose a system that can automatically ingest millions of Online Job Vacancies (OJVs), understand the nuances of skills across different languages and countries, and provide recommendations that reflect real-time market demand rather than static definitions.

Methods and System Architecture

The authors propose a five-step workflow to build the S2JGraph, the backbone of their recommender system:

  1. Linguistic Pattern Extraction: They used FastText to train embedding models on millions of job vacancies. FastText is particularly useful here because it considers sub-word information (character n-grams), which helps the system understand rare words or variations in job titles.
  2. Taxonomic Similarity: To compare skills, they used a measure called Hierarchical Semantic Relatedness (HSS). Unlike standard similarity measures that only look at the "closest common ancestor" in a hierarchy, HSS accounts for the "density" of the taxonomy, making it more sensitive to the specific way skills are grouped in the labor market.
  3. Intrinsic Evaluation: They validated their embeddings by checking how well the cosine similarity between skill vectors correlated with the HSS taxonomic relations.
  4. Skill Relevance: They employed Revealed Comparative Advantage (RCA)—a concept borrowed from economics—to calculate the importance of a specific skill for a specific occupation within a specific country.
  5. Graph Integration: All data (embeddings, RCA scores, and taxonomic relations) is stored in a directed labeled multi-graph database (S2JGraph), which can be queried using the Cypher language.
The system architecture and workflow of the Skills2Job recommender.

Datasets

The system utilizes two primary data sources:

  • Online Job Vacancies (OJVs): A massive corpus of over 2.5 million job postings from the United Kingdom, France, and Germany.
  • ESCO Taxonomy: The European Skills, Competences, Qualifications and Occupations taxonomy was used as the target framework, allowing the system to support 27 different languages.

Key Insights

  • Contextual Skill Importance: A "Web Developer" in the UK might require a slightly different skill set than one in Germany. By calculating RCA and embeddings separately for each country, the system captures these regional nuances.
  • Defining the "Skill Gap": One of the most innovative features is the "skill gap" query. The system identifies skills that are highly relevant to a target job (high RCA) but are significantly different from the user's current skills (low cosine similarity). This helps users identify exactly what they need to learn to level up.
  • Overcoming Taxonomy Limitations: Standard semantic similarity often fails when a word has multiple senses or when a category is too broad. By using HSS, the authors ensure that generic skills don't "drown out" the highly informative, specific skills that actually define a job.

Key Results

The system was evaluated by 10 labor market experts from the European Network on Regional Labour Market Monitoring using a Likert scale. The quantitative results demonstrate high accuracy and ranking performance:

Quantitative evaluation results showing Precision and nDCG.
  • Precision@3-3: 0.823 (meaning over 82% of the top 3 recommendations were considered relevant by experts with a score of at least 3/5).
  • Precision@3-4: 0.610 (relevance score of at least 4/5).
  • nDCG (Normalized Discounted Cumulative Gain): 0.985. This extremely high score indicates that the system is excellent at ranking the most relevant jobs at the very top of the list.

Conclusion and Future Work

Skills2Job demonstrates the power of combining embeddings with graph databases for labor market analysis. While the current implementation focuses on three major European markets, future work aims to scale the system to cover all 26+1 EU countries. Additionally, the authors envision the system as a real-time monitor to help policy-makers observe labor market shifts and skill demands at a granular level across the entire European Union.

Summary updated

BibTeX
@inproceedings{seveso2021skills,
  author = {Andrea Seveso and Anna Giabelli and Lorenzo Malandri and Fabio Mercorio and Mario Mezzanzanica},
  title = {{Skills2Job: A Recommender System that Encodes Job Offer Embeddings on Graph Databases (Student Abstract)}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.1609/aaai.v35i18.17939},
}
Figure