cls2 bug hunt

[ttimbers]

Poorly formatted figures (size, aspect ratio, legibility, etc)

• [ ] Not sure if it can be changed/addressed, but the font in Fig 6.2 is really small at 100% Zoom on my computer. I also wonder if we want to make this figure vertical instead of horizontal to make it easier to fix this? Also, do some of the errors need an explanation of the process, in particular, it is a bit mysterious on the test side where the test data goes from 3 to 2 columns.
• [ ] Can we create more white space between the accuracy equation and Fig 6.2. Also we should make the equations the same perhaps (the one in the text and the one in the figure)?
• [ ] The numbers/data values in Figure 6.4 are very small... Should we try to address this by either maybe only having 3 observations per fold? Or by blurring the numbers intentionally, or referring to them with math symbols (less of a preference due to the audience of the book)?
• [ ] Should we add subplot/subpanel labels to Figure 6.7 to make it easier to refer to them?
• [ ] Figure 6.8 needs a legend for the colors. Also, should validation and test sets be the same color? They are similar but not identical...
• [ ] Should the y-axes in the accuracy graphs start at 0? By not doing that, some of our differences look bigger than they are.
• [x] Figure 6.9 deviates from the formatting of the other similar graphs (accuracy estimate vs neighbours) earlier in this chapter. In particular in the names of the axes, as well as it uses just a line, while the others use a line and a point.
• [ ] For Figure 6.10, the axis has 0.5 increments, but the axis is neighbours... half a neighbour makes no sense...
• [ ] Should Figures 6.10 and 6.11 be combined into a multipanel figure? And if we do so, again should we label them with little letters to make it easier to refer to them?
• [x] Figure 6.10, 6.11 and 6.12 are also just lines, whereas other line charts in the chapter are line + point. Again, we should have a consistent style.

anything else you can think of

• [ ] The more I think about it, the more I think we should number and properly cross reference the equations in the book...
• [x] code style for new stuff related to precision and recall is formatted for PEP8/Black, not tidyverse style. For example, the whites space is this function call should be changed from random_numbers2 <- sample(0:9, 10, replace=TRUE) to random_numbers2 <- sample(0:9, 10, replace = TRUE). There are several examples of this in this chapter, and so it should all be checked over (also happens for the function argument event_level="first"

[ttimbers]

otherwise malfunctioning code

• [x] There is a code output (from knn_fit) under "Evaluating on the test set" that looks like we needs to fix it's width? I only looked in the HTML, maybe we already handle this in the PDF where this is more of an issue?

[ttimbers]

Captions not aligning with figure content

(more like improving the figure captions so that they can more easily stand on their own)

• [ ] Current caption for Figure 6.2 ("Process for splitting the data and finding the prediction accuracy.") is quite terse, and a lot is going on in that figure... We should maybe summarize it in more detail so the figure can better stand on it's own?
• [ ] Should we say we are using standardized data in Figure 6.3?
• [ ] Current caption for Figure 6.4 ("5-fold cross-validation") is quite terse... Similar to Fig 6.2, should we maybe summarize it in more detail so the figure can better stand on it's own?
• [ ] Same terseness issue with Figure 6.8...
• [ ] In figures 6.5, 6.6, 6.9, 6.10, 6.11 and 6.12, should we add more information? Like the model used (KNN) and the data?
• [ ] To be more precise, we should change the figure legend for figure 6.7 from "Effect of K in overfitting and underfitting." to "Effect of K in overfitting and underfitting. Each panel is a scatter plot of concavity versus perimeter colored by diagnosis. The background color indicates the decision of the classifier (which changes dependent upon the value shown for K) and the points represent the labeled data. Too small of a value for K leads to underfitting (e.g., K = 1) and too large a value for K leads to overfitting (e.g., K = 300)." (bolding just to highlight change). Question - again should we be mentioning that this is scaled data?