Healthy Friction in Job Recommender Systems

The paper "Creating Healthy Friction" tackles explainable job recommendation systems from a multi‑stakeholder perspective, recognizing that recruiters, HR teams, and job seekers all influence outcomes. By integrating knowledge‑graph based matching with large language models, the researchers aim to produce human‑readable rationales for each recommendation. This approach addresses a core challenge in recruitment tech: delivering transparency without overwhelming non‑technical users, while also respecting the competing objectives of different parties involved in the hiring process.

In a controlled user study, participants interacted with two explanation styles: authentic, graph‑derived narratives and deliberately random text. Three formats were evaluated—plain textual explanations, graph visualizations, and bar‑chart summaries. Results showed that lay users preferred concise text, finding bar charts confusing and largely ignored. Surprisingly, the gap between real and random explanations was modest; participants used explanations merely as supplemental information rather than decisive evidence, leading to only slight, non‑significant increases in perceived trust and usefulness.

Technically, the system constructs a knowledge graph from candidate and vacancy data, enriches it with inferred relationships, and feeds a JSON representation into an LLM to generate natural‑language explanations. Dual directed graphs capture candidate‑to‑job and job‑to‑candidate perspectives, improving match scoring. The findings suggest that healthy friction—providing explanations without over‑reliance—can enhance user experience while preserving autonomy. Future work should explore larger sample sizes, richer visual aids, and adaptive explanation strategies that dynamically balance stakeholder needs in real‑world recruitment platforms.