New Model Improves Student Performance Prediction in EdTech

Traditional knowledge tracing (KT) models often treat student learning as a single, unified process, which overlooks the distinct phases of learning. This paper highlights that students typically transition from an "ability-building" phase, where they learn new concepts, to a "proficiency-oriented" phase, where they practice and master those concepts. To address this, the researchers propose Phase-Aware Knowledge Tracing (PAKT). This framework decomposes student interactions into these two distinct phases using a tailored mechanism. It then employs a multi-branch Transformer with a type-aware readout module to capture both phase-specific and holistic knowledge states. PAKT's design also includes a causal analysis that reveals and mitigates confounding biases present in phase-agnostic KT models. Extensive experiments across six public benchmarks demonstrate that PAKT consistently outperforms state-of-the-art baselines, showing significant improvements in predicting student performance.