Summary
Endert et al.’s focus in this paper is on how machine learning and visual analytics have blended together to create tools for sensemaking with large complex datasets. They first explain the various models of sensemaking and how they can impact learning and understanding, as well as many models of interactivity in visual analytics that complement sensemaking. Then the authors lightly describe some machine learning models and frameworks to establish a baseline knowledge for the paper. They then create 4 categories for machine learning techniques currently used in visual analytics: dimension reduction, clustering, classification, and regression/correlation models. Then then discuss papers that fit into each of these categories in another set of categories where the user either modifies parameters and computational domain or defines analytical expectations, while the machine learning model assists the user in these. The authors then point out several new ways of blending machine learning and visual analytics, such as steerable machine learning, creating training models from user interaction data, and automated report generation.
Personal Reflection
This paper was an excellent summary of the field of visual analytics and various ways machine learning has been blended into it. Furthermore, there were several papers included that have informed my own research into visual analytics and sensemaking. I was somewhat surprised that, though the authors mention virtual reality, they don’t cover some of the tools that have been developed for immersive analytics. As a side note to this, the authors used a lot of various acronyms and did not explain all of them, for example virtual reality was referenced once and only by its acronym. When they used it for dimensional reduction, I was initially confused because they hadn’t defined that acronym, while they defined the acronyms for visual analytics and machine learning twice in the same paragraph in the introduction.
Their related works section was impressive and really covered a lot of angles for sensemaking and visual analytics. While I do not have the base for machine learning, I assume it also covered that section well.
I also thought the directions they suggested for future development was a good selection of ideas. I could identify ways that many of them could be applied to my work on the Immersive Space to Think, like automated report generation would be a great way to start out in IST, and a way to synthesize and perform topic analysis on any notations while in IST could lead to further analytical goals.
Questions
- What observations do you have on their suggestions for future development in visual analytics? What would you want to tackle first and why?
- In what ways does the human and machine work together in each category of machine learning (dimension reduction, clustering, classification, and regression/correlation)? What affordances does each use?
- Which training method do you think leads to higher quality outputs? Unmodified training sets or user interaction steerable machine learning?
Hi Lee, great reflection. I agree with you that the report did not do a good job of explaining the acronyms. I think part of the reason is that the target audience is the group of researchers who are very familiar with machine learning and visual analytics, which the authors need to improve in order to reach a wider audience. To address your first question, I think their suggestions for future research directions are quite broad and which direction of the first priority may also depend on the need along with the technology development. From my point of view, I think it would be helpful is we can tackle the “Enhancing trust and interpretability” first. The “black-box” design mechanism widely exists in deep learning models limits their usage in a lot of real-world decision-making processes. If we can develop a better understanding of how the models work and their decision boundary, deep learning models would have the potential to be more widely used in our daily life.