Dual-Representation Neural Knowledge Tracing for Auditable AI Assessment in Higher Education

Abstract

This paper propose a Dual-Representation Neural Knowledge Tracing (DR-NKT) architecture as the core engine for auditable AI assessment pipelines in higher education, replacing conventional opaque scoring heuristics with a transparent and interpretable framework. The motivation stems from the urgent need to evaluate student-submitted AI projects—including trained models, scripts, and documentation—while producing a clear reasoning trace for each proficiency score. Our methodology integrates two parallel encoding streams. A Temporal Graph Convolutional Network (TGCN) processes sequential student interactions, such as submission timestamps, autograder logs, and version control metadata, by constructing a temporal graph where nodes represent interaction events and edges capture sequential dependencies through a gated propagation mechanism. Meanwhile, a Symbolic Ontology Parser (SOP) translates formal curriculum standards, derived from educational taxonomies, into differentiable concept vectors using a graph attention network that embeds both semantic content and prerequisite relationships. These two streams are synchronized via a cross-attention fusion module, which computes per-competency mastery scores. Furthermore, the attention weights explicitly link each score to specific interaction events, thereby providing an auditable justification path. The principal contribution is a fully differentiable and auditable assessment submodule that outputs both a skill mastery vector and a set of attention matrices. For example, a low score in “convolutional neural network design” is directly traceable to submissions where the student failed to implement crucial layers. The significance of this work lies in its ability to replace black-box evaluation with transparent reasoning, enabling instructors to audit assessments, provide targeted feedback, and ultimately foster more effective learning in AI education.