Prof. Kobi Gal

We introduce DeepPerfEmb, or DPE, a new deep-learning model that captures dense representations of students’ online behaviour and meta-data about students and educational content. The model uses these representations to predict student performance. We evaluate DPE on standard datasets from the literature, showing superior performance to the state-of-the-art systems in predicting whether or not students will answer a given question correctly. In particular, DPE is unaffected by the cold-start problem which arises when new students come to the system with little to no data available. We also show strong performance of the model when removing students’ histories altogether, relying in part on contextual information about the questions. This strong performance without any information about the learners’ histories demonstrates the high potential of using deep embedded representations of contextual information in educational data mining.