Adaptive Data Visualization via Machine Learning from Human Feedback

Data visualization is a user-centered design process supported by software tools. For example, the software can assist with the decision to use certain visual encodings, such as position, size, and color. The designer can then control parameters that determine precisely where or what size and color each data point will be to meet usability requirements. However, even experts can fail to select effective parameters if they are unaware of unintentional biases that may arise from even the most familiar data representations. This work introduces interactive machine learning tools that help optimize these residual aspects of data visualization design by inferring feedback from data analysts' routine practices. The first contribution incorporates a behavior prediction model that learns from user clicks toward efficiently modeling visual attention given to each data point. The model contributes to the growing literature on visualization biases and can also predict and guide parameter selection for size and color encodings. The second contribution is an AI assistant that utilizes an active preference-based reinforcement learning model to efficiently suggest colormaps for a chart, demonstrating the potential of querying easy-to-input choice data toward efficient visualization parameter selection. Meanwhile, its integration within Jupyter Notebooks allows analysts to stay within their typical workflow and tailor colormaps according to the look-and-feel of the target visualization. The third contribution is an end-to-end qualitative text analysis software where a document clustering model adapts to ongoing user analysis. By inferring feedback from an interpretive researcher's passage highlighting and subsequent 'in vivo' keyword selections, this system seamlessly incorporates expert semantic knowledge into topic modeling and document clustering. Researchers can use the improved interactive corpus visualization to more effectively gather and communicate insights. In summary, I illustrate how intelligent systems that are engineered to ask the right questions can foster effective human-machine collaboration for data visualization by giving and receiving useful feedback.