Characterizing Uncertainty for Sensor Fusion to Improve Predictive Models

Uncertainty is inevitably introduced into machine learning paradigms and threatens the validity and robustness of class prediction. This uncertainty can be introduced at the model level or from the utilized features that are input into the model. This dissertation presents multiple methodologies for evaluating and reducing uncertainty to improve model performance. The work is divided into two sections, which address two types of machine learning problems: handling uncertainty for signature detection (i.e., template matching paradigms) and handling uncertainty for predictive problems for systems with diverse information sources. This work utilizes model fusion methodologies to build upon previously developed machine learning methods to improve the predictive power. The contribution in Part I is a theoretical framework that quantifies and handles model uncertainty assignments allowing multiple sources of information to be fused together in order to improve detection and scalability of temporal and spatial signatures. The contribution in Part II is novel feature engineering techniques and a model fusion approach for improving the downstream effects of uncertainty. Overall, the focus of this work is on the various avenues of how uncertainty enters systems. Through different methods of handling uncertainty, we demonstrate that we can improve the performance of predictive models.