Generating Pseudo-Skill Tags by Extension of Deep Knowledge Tracing

Abstract

Recent progress of EdTech raise attention on the research of the Knowledge Tracing(KT), which tries to estimate the knowledge-level of students by learning log data. On this context, the method so called Deep Knowledge Tracing (DKT), which leverage deep neural network to estimate students’ knowledge-level, shows remarkable performances; however, most of existing KT methods, including DKT, requires skill tags that shows what skill is required to solve a question, and it harms the applicability of KT to real world data, which often not organized by skill tags. In this paper, we extend the DKT model to generate pseudo-skill tags inside the model by enforcing the weight matrix of first layer could be regarded as translation matrix from questions to skills. This paper empirically validate the efficacy of the proposed extension using public two datasets, ASSISTments 2009-2010 and Bridge to Algebra 2006-2007. The results shows our extension gives almost similar or slightly higher performance compared with original DKT model without requiring any pre-defined skill tags. We also analyze the property of generated pseudo-skill tags with statistics and network analysis, and found that they have more hierarchical and information-efficient structure than the pre-defined skill tags.