Computational Assistance – Learning When to Offer Help Precarn/IRIS Student Scholarship Progress Report

In designing computational assistance, one must consider various domain variables as well as individual user differences that influence the effectiveness and transparency of assistance. In this paper, we document a causal analysis under the Bayesian framework for designing computational assistance. In particular, we adopt dynamic Bayesian networks to model the dynamics among the identified factors over time. To demonstrate the approach, we prototyped a word prediction application that interacts with the DBN user model. Unlike most word prediction software, the system does not propose a set of completion words every time the user types in a letter. Rather, it adapts to the style of the user and only makes suggestions when it believes that the user will benefit from the computed suggestions. Such a model is intended to model the styles of different groups of users and adapt to individual preferences. The methodology is generalizable to account for more complex software.