Synthesizing a Knowledge Graph of Data Scientist Job Offers with MINTE+

Mikhail Galkin, Diego Collarana, Maria-Esther Vidal, Mayesha TasnimISWC (P&D/Industry/BlueSky) 2018

Introduction

Analyzing the job market for specialized roles like "Data Scientist" is often a data-wrangling nightmare. Job offers are scattered across diverse platforms—corporate sites, LinkedIn, specialized portals, and community forums—each employing different formats, labels, and levels of granularity. Researchers seeking a holistic view of trending skills or market demands across regions typically face the arduous task of manual scraping, cleaning, and de-duplication.

This paper presents MINTE+, a framework designed to address the challenges of heterogeneous data integration. By treating information as "RDF molecules"—clusters of data centered around a single subject—MINTE+ synthesizes semi-structured data from multiple, potentially conflicting sources into a unified, queryable Knowledge Graph.

Methodology: The RDF Molecule Approach

The core innovation of MINTE+ lies in its molecule-based integration process, which moves beyond flat data structures to preserve semantic context. The framework operates through three primary stages:

  1. RDF Molecule Creation: Instead of processing isolated data points, the system groups "triples" (subject-predicate-object statements) that share a common subject into an RDF Molecule. For a job posting, this molecule encapsulates attributes such as title, company, required skills, and location. The SARO (Skill and Role Ontology) is utilized to provide a standardized vocabulary across these molecules.
  2. Semantically Equivalent Molecule Identification: To determine if postings from different sources (e.g., Indeed vs. Jooble) refer to the same role, MINTE+ employs a Dataset Partitioner to construct a bipartite graph. Similarity is calculated using GADES, a graph-based semantic similarity measure, followed by a 1-1 Weighted Perfect Matching algorithm to pair equivalent molecules.
  3. RDF Molecule Integration: Once matches are identified, an Integrator merges them. The current implementation utilizes a "union fusion policy," combining all unique properties from matched sources into a single master molecule to ensure no information is lost.
The MINTE+ framework architecture, illustrating the pipeline from data ingestion via wrappers to molecule creation, matching, and integration.

Data Sources and Implementation

The framework was demonstrated using live data focused on Data Scientist job offers in Europe from four major sources: Adzuna, Indeed, Jooble, and XING. By leveraging the SARO ontology, the system successfully linked specific technical skills (e.g., Python, TensorFlow) to various job roles, enabling sophisticated semantic queries.

Key Insights

  • Context Preservation: Treating data as molecules rather than flat rows allows the system to handle multi-valued attributes (like a list of five required skills) more naturally, maintaining the context of the original information.
  • Query-Time Synthesis: A significant feature of MINTE+ is its ability to synthesize the knowledge graph at query time. This allows for ad-hoc exploration, where users can adjust integration parameters—such as similarity thresholds—and observe how the resulting graph evolves in real-time.
  • Privacy Considerations: MINTE+ does not provide persistent storage. By serving as a back-end for a faceted browser rather than a permanent database, it mitigates many data protection and GDPR risks associated with the long-term storage of corporate or personal job data.

Experimental Results and Demo

The functional demo of MINTE+ allows users to explore the synthesized job market through a faceted browser. The system effectively de-duplicates identical job postings across platforms. For example, a BMW Group job posting found on both Adzuna and Jooble was correctly identified as a single entity using a similarity threshold of 0.6.

The MINTE+ user interface showing merged job advertisements and a detailed view of an integrated RDF molecule with its corresponding integration log.

Crucially, the system maintains transparency through an "Integration Log," which details the specific ontology, similarity function, and fusion policy used to generate each part of the graph.

Conclusion and Future Work

While MINTE+ successfully demonstrates the feasibility of molecule-based integration, several areas for future development remain. Currently, the "union" fusion policy keeps all data; future iterations could implement more sophisticated conflict-resolution strategies for cases where sources provide contradictory information (e.g., differing salary ranges). Additionally, while the integration itself is automated, the initial "wrappers" for each API were manually implemented. Automating the mapping of new web sources into the RDF molecule format and scaling the graph-matching phase for the global job market are key objectives for future research.

Summary updated

BibTeX
@inproceedings{galkin2018synthesizing,
  author = {Mikhail Galkin and Diego Collarana and Maria-Esther Vidal and Mayesha Tasnim},
  title = {{Synthesizing a Knowledge Graph of Data Scientist Job Offers with MINTE+}},
  booktitle = {ISWC (P\&D/Industry/BlueSky)},
  year = {2018},
}
Figure