Predicting Student Performance Using Sequence Models in XLogoOnline

This study investigates the efficacy of sequence modelling in predicting student performance. Over eight months, user interactions were recorded as students solved navigation tasks in the XLogoOnline programming environment. On this data, three sequence models were trained to learn the temporal dependencies for several performance features. We compared the models' predictive capabilities and found the Transformer architecture to perform best in making multi-step predictions. Although prediction quality declines for multi-step forecasts due to the accumulation of error, we show that the quality of long-term forecast becomes closer to those of shortterm forecasts, as the input length increases. Our results provide valuable insights for the development of more effective teaching tools that can monitor student learning in real time.