Adaptive Learning Trajectory Mapping Using Recurrent Neural Tensor Network in Intelligent Tutoring Systems

Intelligent Tutoring Systems (ITS) have long promised the pedagogical benefits of one-to-one human tutoring at digital scale, but their adaptive capacity is bounded by the quality of their student knowledge models: classical approaches such as Bayesian Knowledge Tracing (BKT) assume a binary mastery state per knowledge component and cannot represent the continuous, multi-dimensional, and temporally evolving nature of genuine student understanding. The Recurrent Neural Tensor Network (RNTN), which extends the LSTM recurrent architecture with a third-order tensor interaction layer that captures multiplicative interactions between the current knowledge state, the incoming exercise response, and the exercise's knowledge component embedding, provides a representationally richer student model that accommodates the gradual, non-linear, and concept-interdependent structure of real learning trajectories. This paper applies RNTN to adaptive learning trajectory mapping across a dataset of 70,545,600 student-exercise interactions from 248,400 learners on three ITS platforms (STEM platform: 84,600 learners; Language platform: 72,400 learners; Mathematics platform: 91,400 learners), spanning 18 subject domains and 48 months (January 2021–December 2024), with a mean of 284 interactions per learner. RNTN achieves an AUC of 0.924 and a response-correctness prediction accuracy of 84.8% on the held-out test set (final 20% of each learner's interaction sequence), outperforming Deep Knowledge Tracing (DKT-LSTM: AUC 0.884), Bayesian Knowledge Tracing (BKT: AUC 0.784), and Item Response Theory (IRT: AUC 0.724). Deployed as the adaptive sequencing engine for all three platforms, the RNTN-driven content recommendation reduces the mean number of interactions required to reach mastery (defined as P(correct) > 0.84 for three consecutive exercises on a knowledge component) by 24.8% compared with static curriculum sequencing, increases 30-day retention to 72.4% (versus 58.4% under static sequencing), and achieves a platform-mean course completion rate of 84.6% versus a 48.4% control group rate.

Hali tarjima qilinmagan