Using multiple Dirichlet distributions to improve parameter plausibility

Predictive accuracy and parameter plausibility are two major desired aspects for a student modeling approach. Knowledge tracing, the most commonly used approach, suffers from local maxima and multiple global maxima. Prior work has shown that using Dirichlet priors improves model parameter plausibility. However, the assumption that all knowledge components are from a single Dirichlet distribution is questionable. To address this problem, this paper presents an approach to integrate multiple distributions and Dirichlet priors. We show that modeling groups of students separately based on their distributional similarities produces model parameters that provide a more plausible picture of student knowledge, even though the proposed solution did not improve the model's predictive accuracy. We also show Dirichlet priors might be hurt by outliers and models with trimming work better.

• Type: Conference
• Year: 2010
• Where: EDM
• Authors: Yue Gong, Joseph E. Beck, Neil T. Heffernan
• Keywords: Data Mining, Dirichlet Priors, EDM 2010, Parameter Plausibility, Predictive Accuracy