Interpretable and Accurate Identification of Job Seekers at Risk of Long-Term Unemployment: Explainable ML-Based Profiling

To tackle the societal and person-specific adverse consequences of long-term unemployment, many public employment services (PES) have implemented data-driven profiling systems to promptly identify vulnerable job seekers. More recently, PES increasingly rely on more complex machine learning (ML) models due to their enhanced accuracy. However, increasing concerns are raised regarding the algorithmic opacity, which hinders comprehension and trust in the predictions. The current study focuses on the explainability of the ML-based profiling model deployed at the Flemish PES (VDAB), aiming to predict clients’ likelihood of securing sustainable employment. We compare two explainability techniques: (1) TreeSHAP is a state-of-the-art method grounded in the theoretical properties of the Shapley values, and (2) TreeInterpreter is a computationally feasible approximation that foregoes some of these properties. Leveraging multiple evaluation metrics, our findings suggest that for tree-based models, approximations to the SHAP (SHapley Additive exPlanations) values yield very similar insights and maintain explanatory performance while minimizing computational overhead. This enables institutions with large client bases to generate real-time explanations without being compelled to deteriorate the model’s accuracy. Additionally, our analysis identifies key predictors of job seekers’ employment prospects, offering valuable insights for PES and related agencies striving to improve their support for job seekers in need. Clients’ online behavior, acting as a proxy for hard-to-measure job search intensity and motivation, emerges as a key component in the profiling model, presenting promising opportunities for future profiling efforts.