|
Abstract:
|
Questionnaires may be used to classify survey participants into different categories in situations where even the participant is unaware of what category they belong to, such as classifying a psychiatric patient into risk categories. We are practically limited in the number of questions we may ask a participant, so we must carefully select the questions to obtain accuracy in classification. Instead of seeking a one-size-fits-all approach, we discuss adaptively selecting questions using information obtained from the participant's previous responses. The presented methodology utilizes ensembles with both model-driven and algorithmic components, supplemented with heuristics and preprocessing of training data to reduce computational burden in order to facilitate near real-time question selection. We discuss prediction of question responses as well as how these predictions assist in selecting future questions. Some complications, such as question order effects are discussed as well. The methods were applied to a simulated data set, and accuracy in classification was then compared to a limited set of benchmarks.
|
