Optimizing Knowledge Tracking and Learning Path Planning Through Virtual Interactions

Knowledge tracking is a task that involves predicting the probability of a student correctly answering a particular exercise. It utilizes the answer interaction data of all students to forecast the answering performance of an individual student. The effectiveness of prediction is derived from the student’s past learning interaction records and the relationships between knowledge concepts embedded in the interaction records of other students. However, the actual volume of data in real learning interaction records is limited, and constrained data may not enable the model to learn the best representation. To address this limitation, we employed the random walk algorithm to generate synthetic data. Although the synthetic data (can be regarded as virtual interaction data) do not exist in reality, they proved effective in training the model, resulting in enhancing performance. The method of generating virtual interactions assist in revealing hierarchical relationships between knowledge concepts that corresponding to exercises, thereby facilitating downstream tasks that require knowledge tracking of predicted results. Experimental results on four classic datasets show that the addition of virtual interaction data can, on average, improve the network’s AUC metric by about 2.2%. Moreover, insights generated from virtual interactions can be applied to learning path planning.