Human–Machine Collaborative Knowledge Tracing for Adaptive and Interactive Physics Laboratory Instruction

Abstract

Accurately characterizing the dynamic evolution of student knowledge states is a foundational prerequisite for designing effective intelligent tutoring systems (ITS) in mobile-supported physics laboratory education. Existing deep learning approaches to knowledge tracing (KT) are predominantly conditioned on student response sequences alone, thereby neglecting two important pedagogical signals in collaborative laboratory environments: human instructor guidance, manifested as teacher demonstrations and corrective interventions, and machine-generated adaptive feedback, manifested as AI-produced hints of varying information quality. To address this limitation, we propose human–machine collaborative knowledge tracing (HMC-KT), a Transformer-based framework that explicitly models the tripartite interaction among student response history, instructor guidance events, and AI feedback quality. Specifically, the human guidance attention (HGA) module encodes instructor intervention signals via gated cross-attention, while the machine feedback attention (MFA) module projects continuous AI hint quality scores into the latent representation space. These two modules are embedded within a causal Transformer backbone and adaptively fused to support context-dependent knowledge state estimation. Extensive experiments on two large-scale public benchmarks—ASSISTments 2009 and ASSISTments 2015—show that HMC-KT achieves consistently superior performance, improving AUC by 3.55 and 4.12 percentage points over the strongest baseline, AKT, respectively, with p < 0.05 under paired t-tests. Systematic ablation studies confirm the independent and complementary contributions of HGA and MFA. Attention visualization further reveals that HGA assigns elevated weights to interaction timesteps following teacher demonstration events, forming a “guidance memory” pattern consistent with the Zone of Proximal Development and Cognitive Load Theory.