Recent advances in Artificial Intelligence (AI) technologies offer exciting opportunities to solve challenging problems with data-driven methods. However, when bringing these technologies built upon machine learning (ML) algorithms from the laboratory to people’s lives, challenges arise from both technical and ethical perspectives. When tackling these issues, a principle is that humans should be put into the center position, i.e., AI empowers and enhances people. Towards this direction, we make visual analytics approaches in response to three progressive, vital questions.

(1) How to provide transparency to ML models: We formulate this problem as probing and explaining the model’s decision boundaries. We explore using counterfactuals (i.e., how to alter a model prediction with minimal ch...[ Read more ]

Recent advances in Artificial Intelligence (AI) technologies offer exciting opportunities to solve challenging problems with data-driven methods. However, when bringing these technologies built upon machine learning (ML) algorithms from the laboratory to people’s lives, challenges arise from both technical and ethical perspectives. When tackling these issues, a principle is that humans should be put into the center position, i.e., AI empowers and enhances people. Towards this direction, we make visual analytics approaches in response to three progressive, vital questions.

(1) How to provide transparency to ML models: We formulate this problem as probing and explaining the model’s decision boundaries. We explore using counterfactuals (i.e., how to alter a model prediction with minimal changes to the data input) to provide truthful and human-friendly explanations. We further develop DECE, a visual analytics system that helps users mentally approximate the model’s decision boundaries by iteratively proposing and refining hypotheses.

(2) How to inform users’ decision-making with explainable ML:We target clinical scenarios and conduct an interview study with the six clinicians to understand the challenges in adopting ML predictions and explanations in clinical decision-making. Following an iterative design process, we further design, develop, and evaluate VBridge, a visual analytics tool that seamlessly incorporates ML explanations into clinicians’ decision-making workflow.

(3) How to incorporate users’ knowledge into ML models: We work with seven molecular biologists to identify the challenges and expectations in applying automatic single-cell annotation tools, which transfer labels from reference datasets (e.g., single-cell atlases) to newly produced data. We further propose Polyphony, a visual analytics system extended from an existing transfer learning method that supports biologists in incorporating their knowledge into the ML model.

This thesis contributes to the fields of visualization, human-computer interaction (HCI), and machine learning with novel interactive visual analytics techniques, design lessons and implications, and open-source software. A list of underexplored directions is further derived from these studies to inspire future research in human-centered AI.