Introduction
Predicting the next step in a professional career—whether for recruitment, upskilling, or workforce planning—is a high-stakes task. However, researchers face a significant "data wall." Most high-quality career data is locked behind the proprietary silos of platforms like LinkedIn or private HR databases.
Publicly available datasets are typically limited to small-scale surveys (often under 15,000 participants) that lack long-term tracking. Furthermore, real-world resumes are written in "free-text," featuring varied and non-standard job titles. Mapping these to a standardized system like ESCO (European Skills, Competences, Qualifications, and Occupations) is notoriously difficult, creating a gap between academic models and real-world application.
Methods
The researchers developed the KARRIEREWEGE dataset using a robust multi-step pipeline designed to bridge the gap between standardized taxonomies and the "messy" nature of real-world resumes.
- Data Sourcing: The process began with anonymized resumes from the German employment agency, covering a wide range of industries.
- Taxonomy Mapping: German resumes were manually mapped to the ESCO taxonomy (version 1.2.0). While automated methods like GPT-4o-mini were tested, human experts were required to ensure high-quality linkages.
- Data Synthesis (KARRIEREWEGE+): To simulate real-world variety, the authors used Llama 3.1 8b to generate paraphrased job titles and descriptions. They compared two strategies: +oc (generating alternative titles for an occupation in isolation) and +cp (generating titles while considering the context of the entire career sequence to ensure logical progression).
- Benchmarking: A state-of-the-art contrastive representation learning model (using
all-mpnet-base-v2) was adapted to predict the next ESCO occupation in a sequence based on the text of previous roles.
Datasets
The project produced two primary resources that significantly expand the available data for career path modeling:
- KARRIEREWEGE: The primary dataset containing 568,888 resumes. It is significantly more diverse than previous benchmarks, covering broader sectors like elementary occupations and service workers.
- KARRIEREWEGE+: A subset of 100,000 resumes enhanced with LLM-synthesized free-text titles and descriptions to better reflect the linguistic variety of actual resumes.
- DECORTE: An existing baseline dataset of ~2,500 English resumes used for comparative benchmarking.
As shown in the distribution comparison, KARRIEREWEGE provides a massive increase in scale, with the frequency of career paths reaching over 160,000 instances compared to the few hundred available in the DECORTE baseline.
Key Insights and Results
The Power of Context and Scale
The researchers found that context is vital for data synthesis; the +cp (career path) method produced higher quality data than the +oc (occupation-only) method. By providing the LLM with the context of a whole career, the generated titles were more coherent and realistic, making the resulting prediction models more robust.
Performance improved dramatically when scaling from 2,000 to 100,000 career paths. However, gains began to taper off after 100,000, suggesting that for this specific task, data diversity eventually becomes more critical than raw volume.
Benchmark Performance
The researchers used Mean Reciprocal Rank (MRR) and Recall@N (R@N) to measure prediction accuracy. Models trained on KARRIEREWEGE consistently outperformed those trained on the DECORTE dataset. On the DECORTE test set, the KARRIEREWEGE-trained model achieved an MRR of 0.4867, nearly doubling the performance of the model trained on the original small dataset (MRR 0.2427).
Limitations and Future Work
While KARRIEREWEGE represents a major step forward, it has limitations. The source data is entirely from the German labor market, which may reflect specific cultural norms and economic structures. Additionally, because the data comes from an employment agency, it may over-represent certain industries or individuals currently seeking work.
Future research directions include expanding the dataset to more regions and languages, and investigating "cross-industry" transitions, which are traditionally much harder to predict than linear promotions within a single field.
Summary updated