Identifying user preferences with Wrapper-based Decision Trees

Wrapper feature selection approaches are widely used to choose a small subset of relevant features from a dataset. However, Wrappers suffer from the fact that they only use a single classifier. The downside to this is that each classifier will have its own biases and will therefore select very different features. To overcome the biases of individual classifiers, we propose a new data mining method called Wrapper-based Decision Trees (WDT). The WDT method uses multiple classifiers for selecting relevant features and decision trees to visualize relationships among the selected features. We use the WDT to investigate the influences of the levels of computer experience on users’ preferences for the design of search engines. The benefit of using WDT lies within the fact that it can uncover the most accurate set of relevant features to help differentiate the preferences of users with diverse levels of computer experience. The results indicate that the users with varied levels of computer experiences have different preferences regarding the following features: the number of icons, the arrangement of search results, and the presentation of error messages. Such findings can be used to develop personalized search engines to accommodate users’ different levels of computer experience.

Research highlights
► 3-classifier combination can identify the most accurate set of relevant features.
► Bayesian Network Classifier leads to low levels of accuracy.
► Average One-Dependence Estimators and Naive Bayes lead to high levels of accuracy.