deletions | additions       diff --git a/subsection_A_Word_About_The__1.tex b/subsection_A_Word_About_The__1.tex  index ef86924..ba33701 100644  --- a/subsection_A_Word_About_The__1.tex  +++ b/subsection_A_Word_About_The__1.tex 
...
\subsection{A Word About The Results}  Stuff. Once the datasets are prepared to be fed to the training process, all the images with the same labeling low/high are lumped together, loosing their provenance from a given user or other. The underlying assumption is that the characteristics of the images should allow us to predict the kind of person that would like them, within the variability (presumed large but not altogether inconsistent) of the users' tastes. The overall accuracy (and associated confusion matrices) tells us how this hypothesis holds up to scrutiny.  In the results reported earlier, it is clear that the self-assessed traits are quite difficult to predict (or the traits values too unreliable, see issue above). The attributed traits instead show a moderate amount of predictability.  After collecting the labels predicted by the classifier, it is possible to breakdown the overall accuracy into per-user accuracy: calculate how accurate are the predictions for the 50 test images of each user separately. The results are shown in the following figure.