Empowering users to communicate their preferences to machine learning models in Visual Analytics
MetadataShow full item record
Recent visual analytic (VA) systems rely on machine learning (ML) to allow users to perform a variety of data analytic tasks, e.g., biologists clustering genome samples, medical practitioners predicting the diagnosis for a new patient, ML practitioners tuning models' hyperparameter settings, etc. These VA systems support interactive construction of models to people (I call them power users) with a diverse set of expertise in ML; from non-experts, to intermediates, to expert ML users. Through my research, I designed and developed VA systems for power users empowering them to communicate their preferences to interactively construct machine learning models for their analytical tasks. In this process, I design algorithms to incorporate user interaction data in machine learning modeling pipelines. Specifically, I deployed and tested (e.g., task completion times, user satisfaction ratings, success rate in finding user-preferred models, model accuracies) two main interaction techniques, multi-model steering, and interactive objective functions to facilitate specification of user goals and objectives to underlying model(s) in VA. However, designing these VA systems for power users poses various challenges, such as addressing diversity in user expertise, metric selection, user modeling to automatically infer preferences, evaluating the success of these systems, etc. Through this work I contribute a set of VA systems that support interactive construction and selection of supervised and unsupervised models using tabular data. In addition, I also present results/findings from a design study of interactive ML in a specific domain with real users and real data.