Interpretable Cognitive State Prediction via Temporal Fuzzy Cognitive Map

Understanding student cognitive states is essential for assessing human learning. The deep neural networks (DNN)-inspired cognitive state prediction method improved prediction performance significantly; however, the lack of explainability with DNNs and the unitary scoring approach fail to reveal the factors influencing human learning. Identifying and understanding these factors remain a challenge. Thus, this article proposes the temporal fuzzy cognitive map (tFCM) model, which combines the prediction power of DNNs with the interpretability of fuzzy cognitive maps. In the proposed tFCM model, cognitive states are modeled as fuzzy, multidimensional, and interrelated vectors, which are input to a long short-term memory network for prediction. This integration allows the proposed model to combine the exceptional ability of DNNs to uncover latent factors with the distinct benefits of fuzzy cognitive maps' ability to reveal potential correlations. A comparative experiment was designed and conducted on a large-scale dataset to assess the predictive performance and interpretability of the proposed tFCM model. The results demonstrate tFCM's superior performance and interpretability compared to existing models. The findings of this study contribute to the development of a multidimensional quantitative model to represent cognitive states and an interpretable model architecture for state prediction.