Towards Job-Transition-Tag Graph for a Better Job Title Representation Learning

Jun Zhu, Céline Hudelot

Abstract

Works on learning job title representation are mainly based on Job-Transition Graph, built from the working history of talents. However, since these records are usually messy, this graph is very sparse, which affects the quality of the learned representation and hinders further analysis. To address this specific issue, we propose to enrich the graph with additional nodes that improve the quality of job title representation. Specifically, we construct Job-Transition-Tag Graph, a heterogeneous graph containing two types of nodes, i.e., job titles and tags (i.e., words related to job responsibilities or functionalities). Along this line, we reformulate job title representation learning as the task of learning node embedding on the Job-Transition-Tag Graph. Experiments on two datasets show the interest of our approach.

Introduction

In HR technology, accurate job title representation is fundamental for downstream tasks like recruitment recommendation and career path analysis. Traditional methods rely on the Job-Transition Graph, where nodes represent job titles and edges represent observed career moves. However, real-world resume data is often noisy and sparse. Variations in naming conventions (e.g., "purchasing clerk" vs. "buyer") and parsing errors lead to a fragmented graph where many semantically similar roles remain disconnected. This sparsity prevents models from learning robust embeddings for less frequent or uniquely named positions.

Methodology

To overcome these limitations, we propose the Job-Transition-Tag Graph (JTTG), a heterogeneous structure that incorporates semantic "tags" as a bridge between job titles.

Tag Extraction and Graph Construction

Tags are extracted from job responsibilities and functionalities. We identify the top 200 most frequent tokens within a domain-specific Named Entity set (e.g., "manager," "technician," "sales"). These tags serve as auxiliary nodes in our graph. We compare four distinct graph formulations:

  1. Job-Transition Graph ($G^{jj}$): The baseline transition-only graph.
  2. Enhanced Job-Transition Graph ($G^{jj}_E$): Adds edges between job titles that share common tags.
  3. Job-Tag Graph ($G^{jt}$): A bipartite graph connecting job titles directly to their constituent tags.
  4. Job-Transition-Tag Graph ($G^{jtj}$): The full heterogeneous graph combining both transitions and tag-based relationships.
Comparison of different job graph structures, showing the evolution from simple transitions to the integrated Job-Transition-Tag Graph.

Representation Learning

We utilize the Heterogeneous Graph Attention Network (HAN) for embedding generation. HAN is particularly suited for this task as its dual-level attention mechanism (node-level and semantic-level) allows the model to weigh the importance of different meta-paths, effectively filtering noise from generic tags while emphasizing defining career transitions.

Experimental Setup

Evaluation was conducted on two datasets:

  • CareerBuilder12 (CB12): A public dataset of job sequences, mapped to standard O*Net-SOC codes for ground-truth classification.
  • Randstad: A private dataset of French resumes, capturing real-world corporate job title diversity.

Results and Discussion

The JTTG approach consistently outperformed transition-only baselines. On the CB12 dataset, the HAN model using the JTTG structure achieved a Macro-F1 of 0.742, a significant improvement over the 0.670 achieved by the standard transition graph. For next-job prediction, the model reached an AUC of 0.833, surpassing other graph-based methods.

Qualitative analysis via t-SNE visualizations confirms that the inclusion of tags helps the model form more cohesive and distinct clusters. As shown in the visualizations, the $G^{jtj}$ structure produces tighter groupings for related occupational categories compared to the sparser $G^{jj}$ baseline.

t-SNE visualizations demonstrating how the Job-Transition-Tag Graph ($G^{jtj}$) produces more distinct and semantically meaningful clusters compared to simpler graph structures.

Conclusion

By enriching job-transition data with semantic tags, we effectively mitigate graph sparsity and improve job title representation learning. Future work will explore more advanced tag extraction methods, such as those leveraging Large Language Models, and the incorporation of hierarchical information like seniority levels into the graph structure.

Summary updated

Abstract

Works on learning job title representation are mainly based on \textit{Job-Transition Graph}, built from the working history of talents. However, since these records are usually messy, this graph is very sparse, which affects the quality of the learned representation and hinders further analysis. To address this specific issue, we propose to enrich the graph with additional nodes that improve the quality of job title representation. Specifically, we construct \textit{Job-Transition-Tag Graph}, a heterogeneous graph containing two types of nodes, i.e., job titles and tags (i.e., words related to job responsibilities or functionalities). Along this line, we reformulate job title representation learning as the task of learning node embedding on the \textit{Job-Transition-Tag Graph}. Experiments on two datasets show the interest of our approach.
BibTeX
@inproceedings{zhu2022towards,
  author = {Jun Zhu and Céline Hudelot},
  title = {{Towards Job-Transition-Tag Graph for a Better Job Title Representation Learning}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2022},
  doi = {10.18653/v1/2022.findings-naacl.164},
}
Figure