Towards Adaptive Ontology Visualization — Predicting User Success from Behavioral Data

Abstract

Listen

Translate article

Ontology visualization plays an important role in human data interaction by offering clarity and insight for complex structured datasets. Recent usability studies of ontology visualization techniques have added to our understanding of desired features when assisting users in the interactive process. However, user behavioral data such as eye gaze and event logs have largely been used as indirect evidence to explain why a user may have carried out certain tasks in a controlled environment, as opposed to direct input that informs the underlying visualization system. Although findings from usability studies have contributed to the refinement of ontology visualizations as a whole, the visualization techniques themselves remain a one-size-fits-all approach, where all users are presented with the same visualizations and interactive features. By contrast, this paper investigates the feasibility of using behavioral data, such as user gaze and event logs, as real-time indicators of how appropriate or effective a given visualization may be for a specific user at a moment in time, which in turn may be used to inform the adaptation of the visualization to the user on the fly. To this end, we apply established predictive modeling techniques in Machine Learning to predict user success using gaze data and event logs. We present a detailed analysis from a controlled experiment and demonstrate such predictions are not only feasible, but can also be significantly better than a baseline classifier during visualization usage. These predictions can then be used to drive the adaptations of visual systems in providing ad hoc visualizations on a per user basis, which in turn may increase individual user success and performance. Furthermore, we demonstrate the prediction performance using several different feature sets, and report on the results generated from several notable classifiers, where a decision tree-based learning model using a boosting algorithm produced the best overall results.