Details

Autor(en) / Beteiligte

• Yu, Hao
• Allessio, Danielle A.
• Rebelsky, William
• Murray, Tom
• Magee, John J.
• Arroyo, Ivon
• Woolf, Beverly P.
• Bargal, Sarah Adel
• Betke, Margrit
• Chounta, Irene-Angelica
• Liu, Zitao
• Santos, Olga C.
• Olney, Andrew M.
• Bittencourt, Ig Ibert

Titel

Affect Behavior Prediction: Using Transformers and Timing Information to Make Early Predictions of Student Exercise Outcome

Ist Teil von

• Artificial Intelligence in Education, p.194-208

Ort / Verlag

Cham: Springer Nature Switzerland

Quelle

Alma/SFX Local Collection

Beschreibungen/Notizen

• Early prediction of student outcomes, as they practice with an intelligent tutoring system is crucial for providing timely and effective interventions to students, potentially improving their engagement and other productive behaviors, and ultimately enhancing their learning. In this work, we propose a novel approach for predicting the outcome of a student solving a mathematics problem in an intelligent tutor using early visual and tabular cues. Our approach analyzes only the first several seconds of a student’s problem-solving process captured in a video feed, along with timing information obtained from their learning log. Our model, EPATT (Early Prediction using an Affect-aware Transformer and Timing information), extracts facial affective embeddings from video frames using transfer learning and analyzes their temporal dependencies using a Transformer. The timing information about when students take certain actions (e.g., requesting a hint) is combined with the video representation, enhancing the model’s ability to predict student performance quickly. Experimental results show that EPATT achieves superior performance over baselines and state-of-the-art on a student dataset for exercise outcome prediction, demonstrating the efficacy of our approach and the potential impact of early outcome prediction for the development of better intelligent tutors.