Deep Explanatory Polytomous Item-Response Model for Predicting Idiosyncratic Affective Ratings

State-Aware Deep Item Response Theory using student facial features

“…Investigating other cultures and ethnicities is an important direction for future research. Delving into individual differences might require more explainable models, potentially at the expense of predictive performance (Rudin, 2019;Zhou et al, 2021), which is another intriguing topic but beyond the scope of the current study.…”

“…To evaluate the interpretability of the proposed model, we compared the estimations of IRT parameters between the proposed model and Base IRT model in the previous study of Zhou et al ( 2021 ). Base IRT model does not have student or item networks, and scalar parameters in the neural network estimate the student and item parameters.…”

“…To explore the ϕ-emotion relationship within our model's framework, we recognize at least two potential methods. One involves employing video-based facial emotion recognizers, as those in Goncalves et al ( 2023 ), for a post hoc analysis, although such analyses are sometimes not advisable (Zhou et al, 2021 ). Alternatively, another classifier layer could be added to our model to process the extracted facial spatiotemporal features.…”

“…Usually, these additional features include items' content information or facial features related to student forgetting behaviors. For example, Zhou et al ( 2021 ) used images as items and trained a Convolutional Neural Network regressor to estimate the item parameter β. This approach demonstrated improved prediction performance compared to methods that used conventional latent regressors to estimate item parameters from handcrafted facial features.…”

“…IRT is based on the local independence assumption that the probability of a person correctly answering a target item does not depend on its response to any other items. In recent years, there have been attempts to combine deep neural networks with IRT to improve prediction performance on responses (Cheng et al, 2019 ; Uto and Uchida, 2020 ; Tsutsumi et al, 2021 ; Zhou et al, 2021 ). The Deep Item Response Theory (Deep-IRT) model (Tsutsumi et al, 2021 ) has shown better response prediction performance than standard IRT models.…”