Machine learning has progressed dramatically in recent decades and become a useful technique in a variety of applications. However, there are various existing machine learning methods, and none of them can work the best for every problem. Moreover, the working mechanism of machine learning models is usually complicated. It is a non-trivial task for practitioners to investigate the behavior of different machine learning models and to select a suitable model for their problems. In this thesis, we design and develop interactive visualization tools, aiming to offer guidance in applying machine learning models.

To help understand machine learning models, we propose DNN Genealogy, a visual summary of representative deep neural networks (DNNs) and their evolutionary relationships. We...[ Read more ]

Machine learning has progressed dramatically in recent decades and become a useful technique in a variety of applications. However, there are various existing machine learning methods, and none of them can work the best for every problem. Moreover, the working mechanism of machine learning models is usually complicated. It is a non-trivial task for practitioners to investigate the behavior of different machine learning models and to select a suitable model for their problems. In this thesis, we design and develop interactive visualization tools, aiming to offer guidance in applying machine learning models.

To help understand machine learning models, we propose DNN Genealogy, a visual summary of representative deep neural networks (DNNs) and their evolutionary relationships. We conduct systematic visualizations of representative DNNs based on our analysis of existing literature. A set of visualization techniques are developed to orient users during their exploration of these DNNs, including a focus-plus-context-based Sugiyamastyle layered graph, a novel performance visualization that combines box plots and bar charts, and a set of network glyphs.

To facilitate the development of machine learning models, we propose an interactive visualization tool to empower an automated model search process. A workflow of using automated machine learning (AutoML) is derived based on expert interviews. Based on this workflow, a multi-granularity visualization is then proposed, aiming to address the lack of transparency and controllability in current AutoML systems.

For a better evaluation of machine learning models, we propose novel visualization techniques and study algorithm discrimination through a visual analytics perspective, which can reveal important details obscured by summary statistics. We identify a collection of potentially discriminatory itemsets based on causal modeling and classification rules mining. By combining an extended Euler diagram with a matrix-based visualization, we develop a novel set visualization technique to facilitate the exploration and interpretation of complex discriminatory itemsets.