Visualizing Uncertainty in Predictive Models - Scientific Visualization: Uncertainty, Multifield, Biomedical, and Scalable Visualization - page 63

Fig. 6.2 Predictions made by model about income level in the census domain, projected using PCA
and displayed with pie glyphs ( left ) and speckle glyphs ( right )
of a set of opaque pixels with density proportional to the class likelihood, and color
corresponding to the class color.
Figure 6.2 shows predictions made by a model constructed to predict income in
the census domain. Based on the demographic attributes, the model predicts whether
an individual will make more or less than $50K a year. The image at the left displays
the predictions using a pie glyph. The size of the white segment in each glyph shows
the predicted probability that the individual makes more than $50K, while the size
of the green segment shows the predicted probability that they make less than $50K.
The ring around the outside of the glyph shows the true class of the instance. Glyphs
for which the ring matches the predominant color of the pie are accurately predicted,
while those with a color mismatch are inaccurate predictions. Notice how most
individuals who make less than $50K a year are predicted with high accuracy, while
those making above $50K a year are much harder to correctly predict. To the upper
center of the image can be seen several fully confident (yet incorrect) predictions.
The right panel of Fig. 6.2 shows the same model displayed using a speckle glyph.
In both pictures, the relative likelihoods of the classes (proportion of low-income
(green) and high-income (white) instances) can be seen fairly clearly. However, the
relationship between the predicted and true class is somewhat more apparent than
with the speckled glyphs, because the density of a class color is easier to visually
interpret as a class probability than the angle width of a region in the pie chart.
Figure 6.3 shows predictions made about a multi-value nominal outcome, specif-
ically, the general occupation of an individual in the census domain. There are 11
occupation values, including tech support, sales, executive/managerial, and machine
operator/inspector. These occupations are predicted based on five input attributes:
age, income, gender, race, and nationality. The images show some localization of pre-
dictions, indicated by grouping of colors, but the regions are much less well defined
than in the income model. In addition, it is apparent that many instances are mis-
predicted (i.e., the most probable predicted class (largest “pie wedge” or dominant
speckle color) is different than the true class), and that there is often high uncertainty
(many different possible classes predicted with approximately equal probability).