In this blog series, international experts on the topic of analyzing online job advertisements present their work. The presentations were given at the 7th OJA Forum, which is jointly organized twice a year by the Federal Institute for Vocational Education and Training and the Bertelsmann Stiftung. Topics at the 7th OJA Forum included the visualization and use of data using dashboards and the assignment of job titles to online job advertisements using various methodological approaches. This article is about the contribution of Kai Krüger from the Federal Institute for Vocational Education and Training. 

Demystifying Deep Learning: Unveiling the Mechanisms Behind a KldB Job Requirement Level Classifier through Explainable AI 

This talk introduces a transformer-based deep learning model designed to predict the fifth digit of the KldB (german classification of occupations), which represents the job requirement level, for a given job posting. Using the explainable AI technique known as Integrated Gradients, we demonstrate which parts of the input—specifically, which words—are most influential in the model’s decision-making process. This analysis helps verify whether the patterns learned by the model represent plausible indicators for classifying the job requirement level, such as job titles, qualifications, or tasks, or if they merely reflect statistical artifacts in the training and evaluation data. 

Further contributions to „7th OJA Forum“:

7th OJA Forum (Part 1) – From data to knowledge on skills

7th OJA Forum (Part 2) – ESCWA’s Sills Monitor

7th OJA Forum (Part 3) – Enhancing the Employability of Students: a LMI Model using OJA

7th OJA Forum (Part 4) – Profession classification in the messy real world

7th OJA Forum (Part 5) – A Hybrid Methodology for Job Ad Title Normalization

7th OJA Forum (Part 6) – Semantic Search with BERT