Jiwon Jason Choi 최지원

If you are korean, check my korean website! https://ko.jasonchoi.dev

We introduce Projection Ensemble, a novel approach for identifying and visualizing robust structures across multidimensional projections. Although multidimensional projections, such as t-Stochastic Neighbor Embedding (t-SNE), have gained popularity, their stochastic nature often leads the user to interpret the structures that arise by chance and make erroneous findings. To overcome this limitation, we present a frequent subgraph mining algorithm and a visualization interface to extract and visualize the consistent structures across multiple projections. We demonstrate that our system not only identifies trustworthy structures but also detects accidental clustering or separation of data points.

We present Intentable, a mixed-initiative caption authoring system that allows the author to steer an automatic caption generation process to reflect their intents on captions. We first derive grammar for specifying the intents, i.e., a caption recipe, and build a neural network that generates caption sentences given a recipe. Our quantitative evaluation revealed that our intent-based generation system not only allows the author to engage in the generation process but also produces more fluent captions than the previous end-to-end approaches without user intervention. Finally, we demonstrate the versatility of our system such as context adaptation, unit conversion, and sentence reordering.

In this paper, we present VANAS, a system that analyzes and visualizes the results of the Neural Architecture Search (NAS) algorithm. First, We devised several algorithms to quantify the significance and contribution of edges, and to recommend neural architecture to users, in order to efficiently convey the vast search results. We created an overview panel to check the overall characteristics of the search space, as well as a user interface to design and analyze the neural network, based on our algorithms. We demonstrated that by analyzing NAS-Bench-101 using VANAS, users can check the importance and significance of nodes and edges that compose neural networks, and that the accuracy can be increased by adding edges recommended by VANAS.

We present Gleaner, an automatic dashboard design system that optimizes the design in terms of four design criteria, namely Specificity, Interestingness, Diversity, and Coverage. With these criteria, Gleaner not only optimizes for the expressiveness and interestingness of a single visualization but also improves the diversity and coverage of the dashboard as a whole. Users are able to express their intent for desired dashboard design to Gleaner, including specifying preferred or constrained attributes and adjusting the weight of each criterion. This flexibility in expressing intent enables Gleaner to design dashboards that are well-aligned with the user’s own analytic goals leading to more efficient data exploration.