Structuring human-ML interaction with an immersive interface based on qualitative coding
Johanne Christensen and Benjamin Watson
Workshop on Immersive Analytics: Exploring Future Interaction and Visualization Technologies for Data Analytics, IEEE Visualization conference (Phoenix, October).
With ever increasing bodies of data, much of it unlabeled and from complex, dynamic and weakly structured domains, machine learning (ML) is more necessary than ever. Yet even domain experts have difficulty understanding most ML algorithms, and so cannot easily retrain them as new data arrives. This limits ML’s use in many fields that sorely need it, such as law, where users must have confidence in ML results. Interactive machine learning techniques have been proposed to take advantage of humanity’s ability to categorize in these complex domains, but little attention has been paid to building interfaces for non-ML experts to provide input, and in particular to creating a user experience that engenders trust. Qualitative coding — the decades-old practice of manual classification — provides a proven methodology that can be adapted to structure interaction between domain experts and ML algorithms. Qualitative
coders often use physical props such as notecards to help sort through and understand datasets. Here we explore how an immersive system can be built to leverage QC’s intuitive techniques and grow a trusting partnership between human and ML classifiers.