This case study illustrates that effectively combining
human intuition and input with machine processing
can enable people to create better classifiers in less
time than standard approaches that ignore these
powerful human capabilities.
Whether a new interface will improve the user’s
experience or the system’s performance can only be
assessed through evaluation with potential end
users. In the case studies above, permitting richer
user interactions was often beneficial, but not always
so. Different users have different needs and expecta-
tions of the systems they employ. In addition, rich
interaction techniques may be appropriate for some
scenarios and not others. Thus, conducting user
studies of novel interactive machine-learning sys-
tems is critical not only for discovering promising
modes of interaction, but also to uncover obstacles
that users may encounter in different scenarios and
unspoken assumptions they might hold about
machine learners. The accumulation of such research
can facilitate the development of design guidelines
for building future interactive machine-learning sys-
tems, much like those that exist for traditional soft-
ware systems (for example, Shneiderman et al.
Interactive machine learning is a potentially power-
ful technique for enabling end-user interaction with
machine learning. As this article illustrates, studying
how people interact with interactive machine-learn-
ing systems and exploring new techniques for
enabling those interactions can result in better user
experiences and more effective machine learners.
However, research in this area has only just begun,
and many opportunities remain to improve the
interactive machine-learning process. This section
Figure 8. EnsembleMatrix.
EnsembleMatrix visualizes the current ensemble (left) of individual learners (bottom right) through a confusion matrix. Users can adjust
the weights of individual models through a linear combination widget (top right) to experiment with different ensembles. Users can also
partition the confusion matrix to split and refine subensembles.
Selected Node Accuracy