Purpose-This study proposes a deep learning parallel dual-classification framework utilizing a lightweight Convolutional Kolmogorov–Arnold Network (C-KAN) to recognize six classes of students' facial expressions and a ResNet-152 to classify three types of teaching methods.

Methods-The parallel framework routes micro-student expressions and macro-classroom contexts separately. The C-KAN model features a two-pronged architecture combining 96× 96 pixel grayscale facial images with 9 geometric features from 68 landmarks, while ResNet-152 processes 256× 256 pixel RGB full-classroom frames.

Findings-Evaluated on elementary school recordings (9,130 teaching method images and 2,069 facial expression instances), ResNet-152 achieved 95% accuracy (0.94 macro F1-score). Meanwhile, C-KAN achieved 87% accuracy (0.86 macro F1-score) across six facial expressions.

Research implications-By leveraging learnable B-spline functions, C-KAN models complex non-linear micro-expressions accurately with only 4.1 million parameters. This structural efficiency slashes inference time to 11 ms, proving its high viability for real-time edge-computing analytics.

Originality-His framework introduces a dual-perspective routing approach. Integrating spline-based C-KAN provides high discriminative power with low computational overhead, supporting evidence-based, automatic evaluation of classroom learning.