Abstract
Matching candidates to roles in hiring, admissions, or volunteer placements is a high-stakes task often plagued by semantic inflexibility, demographic bias, and a lack of transparency. This paper introduces GESA (Graph-Enhanced Semantic Allocation), a multi-tool framework that integrates domain-specific Transformers, heterogeneous Graph Neural Networks, and multi-objective genetic optimization. GESA achieves over 94% accuracy across diverse datasets while significantly improving fairness and providing human-readable explanations for its decisions.
Introduction
Existing automated matching systems frequently fail due to four major flaws: semantic context misses professional nuances, demographic biases are amplified from historical data, "black box" models lack transparency, and it is difficult to balance merit with real-world constraints like diversity quotas. GESA addresses these challenges by modeling the professional landscape as a complex ecosystem rather than a simple list of keywords.
Methodology
The GESA architecture is modular, combining several advanced machine learning techniques to ensure robust performance and ethical alignment:
- IntBERT: A specialized Transformer model fine-tuned on international candidate-role datasets to understand the specific relationships between skills and professional context.
- NexusGNN: A heterogeneous Graph Neural Network that models multi-hop relationships between candidates, roles, skills, organizations, and locations, discovering "indirect pathways" to success.
- Adversarial Debiasing: A built-in discriminator ensures that candidate embeddings do not contain protected attributes (like gender), scrubbing bias at the representation level.
- NSGA-II Optimization: A multi-objective genetic algorithm that simultaneously optimizes for merit, diversity, and preference satisfaction while respecting hard capacity constraints.
- SHAP-based Explainability: Recommendations are accompanied by Shapley value breakdowns, attributing scores to specific features like skill overlap or organizational fit.
Experimental Evaluation
The researchers evaluated GESA across four distinct domains to prove its generalizability: a Synthetic International Dataset (20,000 profiles), an Academic Admissions Dataset (15,000 applicants), a Corporate Hiring Dataset (25,000 applications), and a Volunteer Placement Dataset (8,000 profiles).
Key Results
GESA consistently outperformed baseline models across all metrics. Notably, it achieved a Top-3 Accuracy of 94.5% on the synthetic benchmark, significantly higher than standard GNNs or generic BERT models.
Beyond accuracy, the framework achieved a 0.98 fairness score (near perfect parity) and a 37% improvement in diversity representation compared to merit-only rankings. Furthermore, the system's explainability led to a significantly lower "Override Rate" by human administrators (~8-9% vs. ~25% for keyword matching), indicating higher human trust in the AI's reasoning.
Discussion and Future Work
One of the primary insights of this work is that modeling the professional world as a graph (NexusGNN) provides a significant advantage over looking at candidates in isolation. By understanding the broader ecosystem, the model identifies successful candidates who might be missed by text-only analysis.
While robust, future work will focus on incorporating multi-modal data (such as video interviews), predicting long-term career trajectories, and exploring federated learning to improve models without compromising sensitive private data. Currently, GESA remains a powerful step toward making automated matching both highly accurate and ethically responsible.
Summary updated