Prediction of job careers based on skills using graph kernels

Abstract

This study constructed a model for predicting jobs using the job transition and skills linked to jobs. The model represents each individual as an acyclic undirected graph, which is directly constructed from an actual job transition history. A Weisfeiler-Lehman (WL) graph kernel was then used to construct a feature vector from each graph, and the optimal transport was used to compare them. This approach allows us to incorporate job similarities obtained from a job-skill set dictionary as a cost matrix in optimal transport. Based on this, we evaluated the accuracy of the prediction of the job that an individual with a certain job transition can take next (job node-link prediction). As a result, the proposed model (Skill-based WL, SWL) showed higher accuracy than the model with job transition but no skill set knowledge and those with a mere count of experienced jobs.