Modeling Knowledge Acquisition from Multiple Learning Resource Types

💡 Research Summary

The paper addresses a critical gap in educational data mining: most existing student knowledge tracing models focus on a single type of learning resource—typically problem‑solving data—while modern online learning environments provide a rich mixture of graded (e.g., quizzes, assignments) and non‑graded (e.g., video lectures, discussion posts) materials. Ignoring this heterogeneity limits our ability to accurately model how knowledge is acquired, retained, and transferred across different content formats.

To fill this void, the authors propose the Multi‑View Knowledge Model (MVKM), a unified framework that simultaneously captures student interactions with multiple resource types and discovers latent relationships among the resources themselves. The core idea is to represent each resource type r with a three‑dimensional tensor X⁽ʳ⁾ ∈ ℝ^{M×P^{(r)}×A}, where M is the number of students, P^{(r)} the number of items of type r, and A the sequence of learning attempts (time steps). Each entry x^{(r)}_{a,s,p} records the feedback (grade, binary success/failure, or activity indicator) that student s generated on item p of type r at attempt a.

MVKM factorizes each tensor using a shared low‑dimensional student latent matrix S ∈ ℝ^{M×K}, a time‑varying knowledge tensor T ∈ ℝ^{K×C×A}, and a resource‑concept mapping matrix Q^{(r)} ∈ ℝ^{C×P^{(r)}}. Here K denotes the number of latent student features, C the number of underlying concepts, and A the number of time steps. The predicted score for a graded interaction is:

  \hat{x}_{s,a,p} = s_s · T_a · q_p + b_s + b_p + b_a

where b_s, b_p, and b_a are bias terms accounting for individual ability, item difficulty/helpfulness, and overall course trend, respectively. For non‑graded interactions (or binary‑graded ones) a sigmoid function is applied to the same linear form, yielding a probability of engagement.

A key contribution is the explicit modeling of knowledge growth and occasional forgetting. The authors observe that, under a Markov assumption, knowledge should not decrease after practicing a concept, but real learners sometimes forget. To reconcile these opposing forces, they introduce a rank‑based constraint that penalizes large negative changes in the quantity s_s·T_{a+1}·q_p – s_s·T_a·q_p while still allowing modest declines. This is implemented as an L₂ regularization term on the difference, effectively encouraging smooth, monotonic (but not strictly monotonic) knowledge trajectories.

The “multi‑view” extension shares the student matrix S and the temporal knowledge tensor T across all resource types, while each type retains its own Q^{(r)}. Consequently, knowledge gained from one type (e.g., watching a video) is transferred to the latent space and can influence predictions for another type (e.g., solving a problem). This design directly operationalizes the assumption that learning is transferable across modalities.

Empirical evaluation proceeds on both synthetic data—where ground‑truth latent factors, difficulty levels, and forgetting rates are controlled—and a real‑world dataset collected from an online learning platform containing problem attempts, video view logs, and forum activity. Performance is measured using Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE) on future graded interactions. MVKM consistently outperforms classic Bayesian Knowledge Tracing (BKT), Deep Knowledge Tracing (DKT), Performance Factor Analysis (PFA), and recent multi‑modal baselines. The advantage is most pronounced when non‑graded data are abundant, confirming that auxiliary signals help correct the over‑estimation bias that arises when only graded data are used.

Beyond predictive accuracy, the authors analyze the learned latent representations. By visualizing the Q matrices, they uncover hidden similarities between items of different types that share underlying concepts (e.g., a video explaining a theorem and a problem testing that theorem cluster together). Clustering students based on their S vectors reveals distinct learning‑rate groups (fast vs. slow learners), and the temporal bias b_a captures curriculum pacing effects (e.g., a gradual increase in overall difficulty). These findings demonstrate that MVKM not only predicts performance but also provides interpretable insights for educators and system designers.

In summary, the paper makes four major contributions: (1) a principled multi‑view tensor factorization that jointly models graded and non‑graded interactions; (2) a flexible knowledge‑increase objective that accommodates occasional forgetting; (3) empirical evidence of superior prediction on both synthetic and real data; and (4) a demonstration that the latent factors reveal meaningful cross‑resource relationships and student sub‑populations. MVKM thus offers a comprehensive, scalable solution for personalized learning analytics in heterogeneous online education environments.