Not too long ago, machine learning (ML) methods required significant programming skills in combination with domain expertise and a willingness to manually or programmatically tune the ML system parameters. More sophisticated tools, from front ends to complex ML systems, and eventually automatic ML approaches (AutoML), enabled a wider community to apply ML methods to domain datasets. However, human involvement is still critical for success, and a background in data science, ML, or artificial intelligence (AI) is valuable when it comes to refining a generated system.

The authors describe Visus, an interactive ML model building system intended for domain experts with little or no programming expertise. At the core of the system is an AutoML component that proposes a selection of ML models configured in a variety of data processing pipelines. Interactive visual components assist users with exploratory data analysis, problem specification, the generation and selection of suitable models, and confirmatory data analysis. In addition, data augmentation capabilities support the search for, and integration of, datasets on the web that may cover gaps or expand the set of features. The system also includes limited explanatory capabilities for the suggested configurations and the results obtained.

The proposed system architecture appears sensible, and the benefits for domain experts with limited or no data science and programming skills seem plausible. However, in its current stage, an assessment of the system is difficult. The authors provide a use case scenario and feedback from a small set of users. Once the system is refined and more widely used, it could indeed be beneficial for the intended target audience. While I myself have a background in AI and ML, I would be very interested in a tool like Visus, especially if it offers intuitive visual exploration and validation capabilities for models generated via AutoML.