Self-regulated learning through asynchronous online videos lacks direct social-emotional interactions, often resulting in isolated learning experiences. We present Epimopilot, a system transforming emotional cues from time-synchronized comments (danmaku) into personified virtual companions to create a sense of community. Epimopilot visualizes collective epistemic emotions through emotional companions while providing LLM-powered interventions upon detecting learning difficulties through emotional cues. A within-subjects experiment (N=32) comparing Epimopilot with a baseline system demonstrated significant improvements in learning persistence, engagement, and knowledge comprehension. Results revealed that emotion-infused companion interactions combined with targeted AI assistance significantly enhanced self-regulated learning experiences. The analysis yielded design principles for recognizing epistemic emotions and integrating collective emotional cues with a knowledge graph. This work contributes an innovative approach to foster a co-presence learning experience in asynchronous environments through emotion recognition and adaptive cognitive support.

Knowledge tracing is a foundational task in intelligent education, aiming to predict students’ future performance by modeling their historical interactions. Traditional knowledge tracing methods primarily focus on analyzing students’ learning behaviors, often neglecting their dynamically changing emotional states during the learning process. Given that emotion is an integral part of the cognitive process, neglecting emotional states will limit the representation capacity of knowledge tracing models for modeling students’ learning processes. To address this issue, we propose a novel emotion-aware knowledge tracing method to explore the impact of complex emotions (e.g., concentration, frustration, boredom, and confusion) on knowledge states throughout the learning process. Specifically, we design an emotion-aware fusion module to capture the joint influence of multiple coexisting emotional states on students’ answers during each exercise session. Additionally, we utilize graph convolutional networks to propagate embeddings and obtain the complete knowledge structure information of exercises. We further integrate exercise attributes and students’ emotional information through a feature fusion module, thereby incorporating emotional states into the knowledge tracing task. Finally, we apply a transformer-based knowledge evolution module to model students’ evolving knowledge states, achieving multiview fusion modeling of emotions, knowledge topology, and learning sequences. Extensive experiments demonstrate that our method outperforms previous knowledge tracing methods in predicting student performance.