GV: EAGER: Innovative Analysis and Visualization Approaches for Understanding Model Uncertainty

点击翻译按钮获取中文摘要

This exploratory project strives to develop a new approach to support human understanding of the uncertainty that is inherent in the structure and predictions of complex models.Specifically, the focus in the project is on understanding several types of uncertainty that are associated with model predictions.Sample uncertainty occurs when regions of the instance space are not well represented in the training data, and predictions are therefore based on sparse information. Model instability occurs when model predictions vary, depending on the training data that was used to construct the model. Prediction variability occurs when a given observation may have noisy attributes, and this input uncertainty leads to uncertainty in the model's predictions. Novel analytical techniques are developed to create meta-models that characterize these three forms of uncertainty. To facilitate user understanding of the nature and distribution of these multiple types of uncertainty across the model space, novel visualization methods represent these meta-models in a display space. Finally, a novel evaluation methodology is used to measure whether, and in what ways, important characteristics of the meta-models are captured in the visualization display space.This work develops novel techniques in the fields of machine learning and data visualization. Contributions in machine learning include more powerful methods for constructing and analyzing meta-models that characterize multiple types of uncertainty associated with predictive models. Data visualization research focuses on new approaches for representing multi-valued, probabilistic, and complex data, enabling the display of the nature and range of model predictions and uncertainty. An interdisciplinary contribution is the development of a novel methodology for evaluating the quality of model visualizations with respect to the preservation of important model and meta-model characteristics.The broader impacts of this project may be grouped into three major clusters: a new model building paradigm; fostering scientific collaboration; and integrating research and education. The results are expected to provide foundations for further research is management of uncertainty in deriving models representing a wide range of phenomena. This project lays a technical groundwork that can contribute to new collaborations between the PIs and application domain experts, facilitating broad interdisciplinary collaborations. Project results will be widely disseminated via the project web site (http://maple.cs.umbc.edu/complexmodels/). Finally, through teaching and training activities, this research project is also well suited to include the introduction of undergraduates to the possibilities of research and the incorporation of project topics into the PIs' courses on visualization and artificial intelligence.