As the plots in ggdibbler are random variables, they are
random plots, which means each time you call a plot, it is different.
This behaviour makes sense, as every plot is just one of many possible
outcomes you can have when you represent a distribution by a sample. If
you are used to the deterministic behaviour of ggplot, the randomness
can be off-putting, although not technically incorrect. Most of these
issue can be mitigated by using the seed option, but by default the
plots are random.
When you save a plot, ggplot2doesn’t technically save
the plot itself. It saves the code that makes the plot. It only actually
runs the code when ggplot_build() is called, which is only
called when you print the plot.
The two scatter plots above were called using the same variable
plot and would be identical if they were a ggplot, but you
can see they are slightly different.
ggdibbler assumes all aesthetic distributions passed are
independent, which means if you pass the same variable to two different
aesthetic, they will be different samples.
p1 <- ggplot(mtcars, aes(wt, wt)) + geom_point() +
ggtitle("ggplot2")
p2 <- ggplot(uncertain_mtcars, aes(wt, wt)) + geom_point_sample(times=5) +
ggtitle("ggdibbler")
p1 + p2
By the time a column gets to the stat where we draw the samples, we
do not have the variable names any more. This means we can’t be sure
which variables are actually identical versus which variables only
appear to be identical. You can get the identical behaviour by one of
the draws using the after_stat function.
This problem is particularly problematic when you have multiple
layers using the same variable where the after_stat trick
no longer works. In this case so you will need to use the seed parameter
to ensure the draws are the same.
recent <- economics[economics$date > as.Date("2013-01-01"), ]
uncertain_recent <- uncertain_economics[uncertain_economics$date > as.Date("2013-01-01"), ]
p1 <- ggplot(recent, aes(date, unemploy)) +
geom_line() +
geom_point() +
ggtitle("ggplot2")
p2 <- ggplot(uncertain_recent, aes(date, unemploy)) +
geom_line_sample(alpha=0.5, seed=5)+
geom_point_sample(alpha=0.5, seed=5) +
ggtitle("ggdibbler")
p1 + p2
In showing ggdibbler to people, I have noticed that
there are a few instances where the package does not behave in the way
they expect. Sometimes this is because of a bug on the packages part,
but I have also found that it can be because people are unsure what they
mean when they talk about uncertainty. ggdibbler will
always show you the plot you asked for, but it might not always be the
plot you think you asked for. This will be easiest to see with an
example.
Let’s say we are looking at the smaller_diamonds dataset
(which is just the diamonds dataset from
ggplot2, but… smaller) and we want to predict the cut using
the other variables. We make our fancy shmancy model, and then get a
predicted value for each class. Since our output is a prediction, we now
have a cut_true column, (our ground truth) and a
cut_pred which is our predicted distribution.
## # A tibble: 6 × 11
## cut_pred cut_true carat color clarity depth
## <dist> <ord> <dbl> <ord> <ord> <dbl>
## 1 Categorical[5] Very Good 0.55 F VS2 62.1
## 2 Categorical[5] Premium 2 G SI2 61.7
## 3 Categorical[5] Ideal 0.31 F VVS2 61.6
## 4 Categorical[5] Premium 1.52 I SI1 60.5
## 5 Categorical[5] Good 1.01 G VS2 62
## 6 Categorical[5] Ideal 0.82 I VS1 61.3
## # ℹ 5 more variables: table <dbl>, price <int>, x <dbl>, y <dbl>, z <dbl>You might want to know how many values end up in each group and how certain the model is in that prediction, so, you make the following bar chart:
Looking at this plot, it is clear that the count value is uncertain, but
that uncertainty is coming from the uncertainty in the cut_prediction.
The cut_pred values don’t look very “uncertain”, because in this plot,
they are not. This is the reality of visualising uncertainty for signal
suppression. Since the stat_sample that is used to get
outcomes from the distribution is nested inside of ggplot’s
stat_count, the uncertainty in the prediction is carried
through to the variable the plot is designed to show, i.e. the count in
each category. If it makes you feel better, we can set the fill to be
the outputs so you can 100% see this is the case.
ggplot(diamonds_pred, aes(x=cut_pred)) +
geom_bar_sample(aes(fill=factor(after_stat(x)))) +
labs(fill = "cut_pred")+
scale_fill_brewer(palette = "Set1")
Of course, you can also set the fill to be a second independent draw from the prediction distribution, but this would convey the wrong information. You would be implying far more uncertainty than is actually in the bar chart. Instead of drawing from one distribution, we are drawing from two independent distributions that technically represent the same variable.
The reality is, distribution variables are “slippery” in a way that deterministic variables are not. They do not have a set value, so if their position on the x axis and their colour are not set by the sample outcome…. what are they set by? The reality, is, if you want the random variables anchored to something, YOU need to decide what to anchor them to. You need to decide what your random variables are conditional on. If you want them conditional on the deterministic true value, you need to do that. Here is an example where they are anchored to the ground truth value and we colour by the prediction.
p1 <- ggplot(diamonds_pred, aes(x=cut_true)) +
geom_bar() +
ggtitle("No colour randomness")
p2 <- ggplot(diamonds_pred, aes(x=cut_true)) +
geom_bar_sample(aes(fill= cut_pred), times=100,
position= "stack_dodge") +
scale_fill_brewer(palette = "Set1") +
ggtitle("Random colour (dodge)")
p3 <- ggplot(diamonds_pred, aes(x=cut_true)) +
geom_bar_sample(aes(fill= cut_pred), times=100, alpha=0.1,
position= "stack_identity") +
scale_fill_brewer(palette = "Set1") +
ggtitle("Random colour (identity)")
p2 | p3
The height of the bar represents the total number of values predicted as that class, and the proportion of the the bar that is filled in as that colour represents the probability that the observation is that class. You will notice that the y-axis now represents the count of a deterministic variable (so there is no random variation and it is easy to read) but the colour is random, so it does have variation.