A Generative Approach for Proactive Assistance Forecasting in Intelligent Tutoring Environments

Adaptive user feedback facilitates the delivery of personalized assistance to students struggling with self-learning and enhances their overall learning effectiveness. However, numerous studies on student behavior have revealed that they may not consistently utilize help-seeking functions. Deciding when a system should assist students during the dynamic learning process poses a challenge. We propose a new approach called Transformer4HELP, which enables the system to proactively assist students in their learning process interactions in a generative way. We employ an auto-regressive masking strategy to train a decoder-based transformer for prediction, assessing the necessity of providing help at each time step based on the likelihood that intervention may benefit the user’s actions. To evaluate this approach, we used real behavioral data from students engaged in solving arithmetic mathematical problems. The experimental results demonstrate the effectiveness of the proposed method, reaching an AUC of 0.84 when predicting whether the student needs help.