Techniques for Combining Visualization with Machine Learning in Data Analysis

Abstract

In recent years, visualization techniques and machine learning have revolutionized the way people understand data. While interactive visual systems allow the user to summarize the data and identify discriminative features or outliers, machine learning techniques train predictive models to discover naturally occurring patterns in the datasets. There has been increasing interest in utilizing the advantages of visualization and machine learning concurrently for data analysis. However, the extant state techniques still have limitations, such as low model interpretability. In this dissertation, I present two techniques, STFT-LDA and ConceptExtract, combining machine learning methods with visual analytics systems to understand complicated datasets in different domains. STFT-LDA significantly improves civil engineers' understanding of simulations of seismic responses by utilizing a topic modeling method to extract interpretable and discriminative features of the earthquakes. ConceptExtract leverages a novel human-in-the-loop approach to generate user-defined concepts to help interpret start-of-art neural networks. The learned concepts serve as important clues for model diagnostics and improvements. Machine learning approaches are carefully integrated with visual analytics systems through these two projects, demonstrating how this combination can simplify visual encoding and save time and human effort to discover interesting patterns.