Cognitive-Sensor-Aware Home Learning Control: Personalized Multimodal Adaptation Via Meta-Learning

This study proposes a cnovel ognitive-aware home learning control model. This model implements a groundbreaking meta-learning approach for the multimodal personalized adaptation. The proposed model uses multimodal sensors to collect learners' emotional, cognitive, psychophysiological, and environmental data in real time. It also constructs a high-dimensional tensor joint representation to capture multimodal interaction characteristics. To address the noise and uncertainty of sensor signals, this study introduces a novel dynamic Bayesian network for state transition modeling. Then, information theory and Wasserstein-1 distance are combined as optimization tools to quantify modal contributions and optimize personalized control objectives. To address the differences in modal dependence among learners, this paper obtains rapid adaptive adjustment of personalized sensor weights through dynamic factor modeling and a meta-learning strategy. Results from a short-term experiment (20 middle school students, 30 minutes each session) showed that the proposed method improved the Pearson correlation coefficient by approximately <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{1 5 \% - 2 5 \%}$</tex> and reduced the RMSE by approximately <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{3 0 \%}-\mathbf{4 0 \%}$</tex> compared to the uni-modal approach. Long-term experiments (7 days, <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{1}$</tex> hour per day, totaling approximately <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{8 4 0}$</tex> hours of data) showed that the predictive correlation of the uni-modal approach decreased significantly, while the proposed method remained stable. Numerically, the Pearson correlation coefficient decreased only slightly by approximately 2 %, while the RMSE remained essentially unchanged. The modal weights can be dynamically adjusted to adapt to long-term learning changes.