Per-pitch BKT attains the highest accuracy, lowest BCE and lowest Brier score of the four model families, but lower AUC than IRT¶

Claim · All claims
Evidence · 1 study · associational r2 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims¶

q2 i? BKT achieved accuracy 0.7520, BCE 0.5528 and Brier 0.180, beating IRT on accuracy by 0.025, while its AUC of 0.653 was 0.019 below the IRT AUC. → Wei 2026

Evidence¶

Wei 2026¶

Wei, Wang and Dong. (2026). Cognitive and skill acquisition trajectories in school-based music education: evidence from Chinese classrooms. Frontiers in Psychology. https://doi.org/10.3389/fpsyg.2026.1905847

q2 · i? · associational · r2

Held-out evaluation (Table 3) on the identical 492,633-attempt test partition; BKT parameters were estimated per pitch by forward–backward expectation–maximization. Both pairwise differences had bootstrap intervals excluding zero.

"The BKT model achieved the lowest test BCE [0.5528 (0.534, 0.574)], the highest accuracy [0.7520 (0.740, 0.763)] and the lowest Brier score [0.180 (0.174, 0.188)] of all four families, with an AUC of 0.653 [(0.642, 0.664)] that is lower than the IRT AUC"

Discussion¶

Related Claims¶

• The survey states that deep learning knowledge tracing models exhibit superior performance to Bayesian and logistic models but have significant room for improvement in interpretability. — the general claim this finding bears on, with the studies behind it
• A one-parameter IRT model outperforms global and per-pitch baselines on AUC, BCE and Brier score but not on threshold accuracy for held-out singing attempts — related
• BKT is the most efficient knowledge tracing approach overall in simulated online mastery learning, though DKT is more efficient for AS and M problems — related
• Low-register pitches are sung more accurately than high-register pitches in Chinese classrooms, reproducing register compression at scale — related