Abstract
Field Computer-implemented systems for predicting employee knowledge deficits by analyzing the structural topology of an organizational skill dependency graph using graph neural networks. Background Traditional approaches to detecting employee knowledge gaps rely on analyzing what employees say (natural language processing) or how they score on assessments (threshold comparison). The approach described here predicts gaps from the structure of relationships between employees, skills, and tasks without processing any communications or administering any assessments. Technical Description The system constructs a multi-partite graph with three node types: employee nodes E, skill nodes S, and task nodes T. Edges connect employees to skills they have demonstrated (proficiency edges), skills to tasks that require them (requirement edges), and skills to other skills (prerequisite edges). Edge weights represent confidence levels derived from assessment scores, task completion rates, and manager certifications. A Graph Convolutional Network learns latent representations for all nodes using a message-passing framework. At each layer k, each node v updates its embedding by aggregating information from its neighborhood: h_v at layer k equals sigma of W at layer k multiplied by the aggregation of h_u at layer k minus 1 for all u in the neighborhood of v. The aggregation function can be mean, sum, or attention-weighted. The GCN is trained on historical data where known proficiency edges serve as positive examples and known gaps serve as negative examples. Gap detection operates through link prediction. For each employee-skill pair where no proficiency edge currently exists, the system computes a predicted probability that the edge should exist based on the learned embeddings. Specifically, P(E_i, S_j) equals the sigmoid of the dot product of h(E_i) and h(S_j). When this probability exceeds a configurable threshold (default: 0.7) but the actual edge is absent, the system generates a structural gap record. A constraint violation detector provides a second detection pathway. When an employee has a proficiency edge to a higher-order skill but lacks an edge to a prerequisite skill, a violation score is computed: violation = mastery(higher_skill) times (1 minus mastery(prerequisite_skill)). Violations exceeding a threshold generate gap records. The gap record schema contains: gap_id, employee_id, skill_id, prediction_probability or violation_score, detection_method (link_prediction or constraint_violation), dependent_task_ids, and timestamp. Distinguishing Characteristics This system does not process inbound natural language communications. No NLP pipeline, no intent classification, no entity extraction, no taxonomy mapping from text. It does not use structured assessment scores or threshold comparisons against quiz results. Gap detection comes entirely from the mathematical properties of graph connectivity and learned node embeddings. The gap signal is a missing or predicted edge in the skill graph, not a classified communication event or a score belo
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Davis, Kenneth, "Graph Neural Network Structural Skill Gap Prediction via Link Prediction on Employee-Skill-Task Graphs", Technical Disclosure Commons, (March 25, 2026)
https://www.tdcommons.org/dpubs_series/9628