RA4KT: a causality-informed explainable reverse attribution framework for knowledge tracing

With the global digital transformation of education, Knowledge Tracing, the core cognitive modeling module in intelligent tutoring systems, has become a key research focus in educational data mining. However, existing deep learning-based KT models are inherently black-box: while achieving high-accuracy answer performance prediction, they fail to clarify internal decision logic or translate outputs into actionable teaching guidance for students and teachers. Existing KT interpretability methods cannot simultaneously satisfy the core requirements of causality, actionability, pedagogical adaptability, and non-professional accessibility, severely limiting KT’s industrial deployment and compliant application of educational AI systems. To address these research gaps, we systematically establish the conceptual definition, mathematical formalization, and five core attributes of reverse attribution for KT sequence learning scenarios, and propose RA4KT, a model-agnostic post-hoc explainable framework specifically tailored to the sequential characteristics and pedagogical prior constraints of educational scenarios. RA4KT unifies attribution validity, modification minimalism, and pedagogical coherence via a multi-objective loss function integrating knowledge concept prior correlations, enforces intervention actionability through an Actionability Mask hard constraint mechanism, and connects model explanation to teaching intervention via a Traveling Salesman Problem (TSP)-variant post-processing module that generates cognitively aligned ordered teaching paths. Systematic experiments on two large-scale real-world educational datasets (XES3G5M and ASSISTments2017) show that RA4KT outperforms mainstream reverse attribution baselines (DiCE, CETD, CMKT, RCKT) on core metrics (Validity, Sparsity, Pedagogical Relevance, Robustness), with relative improvements of the best-performing RA4KT-gs variant over the strongest baseline on each core metric ranging from 2.1% to 16.6% (per-metric, per-dataset values and calculation details in Tables 2–3 and Section "Main experimental results and analysis"), maintains stable performance in long-sequence scenarios, and delivers significant knowledge mastery efficiency gains in offline intervention simulations, especially for students with low initial mastery levels. This work fills the technical gap between black-box KT prediction and actionable intervention schemes, providing a technical pathway toward interpretable engineering deployment of educational AI, with potential extensibility to high-stakes time-series scenarios including healthcare, financial risk control, and industrial predictive maintenance, subject to domain-specific validation.