Constructing an Adaptive Training Model for Electronic Information Postgraduates Using Causally Interpretable Artificial Intelligence

Abstract

The cultivation of innovation capability in electronic information postgraduates faces multiple obstacles, including complex and nonlinear interactions among influencing factors and consistently delayed evaluation feedback. As a result, conventional training models cannot adapt dynamically to individual characteristics. To address this limitation, this paper proposes an adaptive training model grounded in causally interpretable artificial intelligence. First, a three-dimensional feature set encompassing individual endowments, training process support, and the external ecological environment is constructed. Subsequently, a neural acyclic directed mixed graph learning method extracts the causal network structure among these features, identifying critical pathways and potential confounders. Finally, this causal structure is incorporated as prior information into an XGBoost-SHAP framework, which produces a dual-output evaluation model that simultaneously generates personalized student innovation capability reports and supervisor effectiveness reports. The model achieves adaptive evolution through a dual closed-loop mechanism, in which causal discovery drives iterative feature set updates and evaluation feedback drives dynamic adjustments to training strategies. This research offers a new pathway that fuses causal mechanisms with interpretable artificial intelligence for cultivating innovation capability in electronic information postgraduates.