22
$\begingroup$

Among others on here, Frank Harrell is adamant about using proper scoring rules to assess classifiers. This makes sense. If we have 500 $0$s with $P(1)\in[0.45, 0.49]$ and 500 $1$s with $P(1)\in[0.51, 0.55]$, we can get a perfect classifier by setting our threshold at $0.50$. However, is that really a better classifier than one that gives the $0$s all $P(1)\in[0.05, 0.07]$ and the $1$s all $P(1)\in[0.93,0.95]$, except for one that has $P(1)=0.04?$

Brier score says that the second classifier crushes the first, even though the second cannot achieve perfect accuracy.

set.seed(2020)
N <- 500
spam_1 <- runif(N, 0.45, 0.49) # category 0
ham_1 <- runif(N, 0.51, 0.55) # category 1
brier_score_1 <- sum((spam_1)^2) + sum((ham_1-1)^2)
spam_2 <- runif(N, 0.05, 0.07) # category 0 
ham_2 <- c(0.04, runif(N-1, 0.93, 0.95)) # category 1
brier_score_2 <- sum((spam_2)^2) + sum((ham_2-1)^2)
brier_score_1 # turns out to be 221.3765
brier_score_2 # turns out to be 4.550592

However, if we go with the second classifier, we end up calling a "ham" email "spam" and sending it to the spam folder. Depending on the email content, that could be quite bad news. With the first classifier, if we use a threshold of $0.50$, we always classify the spam as spam and the ham as ham. The second classifier has no threshold that can give the perfect classification accuracy that would be so wonderful for email filtering.

I concede that I don't know the inner workings of a spam filter, but I suspect there's a hard decision made to send an email to the spam folder or let it through to the inbox.$^{\dagger}$ Even if this is not how the particular example of email filtering works, there are situations where decisions have to be made.

As the user of a classifier who has to make a decision, what is the benefit of using a proper scoring rule as opposed to finding the optimal threshold and then assessing the performance when we classify according to that threshold? Sure, we may value sensitivity or specificity instead of just accuracy, but we don't get any of those from a proper scoring rule. I can imagine the following conversation with a manager.

Me: "So I propose that we use the second model, because of its much lower Brier score."

Boss: "So you want to go with the model that [goofs] up more often? SECURITY!"

I can see an argument that the model with the lower Brier score (good) but lower accuracy (bad) might be expected to perform better (in terms of classification accuracy) in the long run and should not be so harshly penalized because of a fluke point that the other model gets despite its generally worse performance, but that still feels like an unsatisfying answer to give a manager if we’re doing out-of-sample testing and seeing how these models perform on data to which they were not exposed during training.

$^{\dagger}$An alternative would be some kind of dice roll based on the probability determined by the classifier. Say we get $P(spam)=0.23$. Then draw an observation $X$ from $\text{Bernoulli}(0.23)$ and send it to the spam folder iff $X=1$. At some point, however, there is a decision made about where to send the email, no "23% send it to the spam folder, 77% let it through to the inbox".

$\endgroup$
1
  • 2
    $\begingroup$ If you don't care for probabilities or uncertainty, then ditch it altogether and slap a hinge loss on your problem instead. $\endgroup$
    – Firebug
    Jun 16 '20 at 20:12
18
$\begingroup$

I guess I'm one of the "among others", so I'll chime in.

The short version: I'm afraid your example is a bit of a straw man, and I don't think we can learn a lot from it.

In the first case, yes, you can threshold your predictions at 0.50 to get a perfect classification. True. But we also see that your model is actually rather poor. Take item #127 in the spam group, and compare it to item #484 in the ham group. They have predicted probabilities of being spam of 0.49 and 0.51. (That's because I picked the largest prediction in the spam and the smallest prediction in the ham group.)

That is, for the model they are almost indistinguishable in terms of their likelihood of being spam. But they aren't! We know that the first one is practically certain to be spam, and the second one to be ham. "Practically certain" as in "we observed 1000 instances, and the cutoff always worked". Saying that the two instances are practically equally likely to be spam is a clear indication that our model doesn't really know what it is doing.

Thus, in the present case, the conversation should not be whether we should go with model 1 or with model 2, or whether we should decide between the two models based on accuracy or on the Brier score. Rather, we should be feeding both models' predictions to any standard third model, such as a standard logistic regression. This will transform the predictions from model 1 to extremely confident predictions that are essentially 0 and 1 and thus reflect the structure in the data much better. The Brier score of this meta-model will be much lower, on the order of zero. And in the same way, the predictions from model 2 will be transformed into predictions that are almost as good, but a little worse - with a Brier score that is somewhat higher. Now, the Brier score of the two meta-models will correctly reflect that the one based on (meta-)model 1 should be preferred.


And of course, the final decision will likely need to use some kind of threshold. Depending on the costs of type I and II errors, the cost-optimal threshold might well be different from 0.5 (except, of course, in the present example). After all, as you write, it may be much more costly to misclassify ham as spam than vice versa. But as I write elsewhere, a cost optimal decision might also well include more than one threshold! Quite possibly, a very low predicted spam probability might have the mail sent to your inbox directly, while a very high predicted probability might have it filtered at the mail server without you ever seeing it - but probabilities in between might mean that a [SUSPECTED SPAM] might be inserted in the subject, and the mail would still be sent to your inbox. Accuracy as an evaluation measure fails here, unless we start looking at separate accuracy for the multiple buckets, but in the end, all the "in between" mails will be classified as one or the other, and shouldn't they have been sent to the correct bucket in the first place? Proper scoring rules, on the other hand, can help you calibrate your probabilistic predictions.


To be honest, I don't think deterministic examples like the one you give here are very useful. If we know what is happening, then we wouldn't be doing probabilistic classification/prediction in the first place, after all. So I would argue for probabilistic examples. Here is one such. I'll generate 1,000 true underlying probabilities as uniformly distributed on $[0,1]$, then generate actuals according to this probability. Now we don't have the perfect separation that I'm arguing fogs up the example above.

set.seed(2020)
nn <- 1000
true_probabilities <- runif(nn)
actuals <- runif(nn)<true_probabilities

library(beanplot)
beanplot(true_probabilities~actuals, 
    horizontal=TRUE,what=c(0,1,0,0),border=NA,col="lightgray",las=1,
    xlab="True probability")
points(true_probabilities,actuals+1+runif(nn,-0.3,0.3),pch=19,cex=0.6)

beanplot

Now, if we have the true probabilities, we can use cost-based thresholds as above. But typically, we will not know these true probabilities, but we may need to decide between competing models that each output such probabilities. I would argue that searching for a model that gets as close as possible to these true probabilities is worthwhile, because, for instance, if we have a biased understanding of the true probabilities, any resources we invest in changing the process (e.g., in medical applications: screening, inoculation, propagating lifestyle changes, ...) or in understanding it better may be misallocated. Put differently: working with accuracy and a threshold means that we don't care at all whether we predict a probability $\hat{p}_1$ or $\hat{p}_2$ as long as it's above the threshold, $\hat{p}_i>t$ (and vice versa below $t$), so we have zero incentive in understanding and investigating which instances we are unsure about, just as long as we get them to the correct side of the threshold.

Let's look at a couple of miscalibrated predicted probabilities. Specifically, for the true probabilities $p$, we can look at power transforms $\hat{p}_x:=p^x$ for some exponent $x>0$. This is a monotone transformation, so any thresholds we would like to use based on $p$ can also be transformed for use with $\hat{p}_x$. Or, starting from $\hat{p}_x$ and not knowing $p$, we can optimize thresholds $\hat{t}_x$ to get the exact same accuracies for $(\hat{p}_x,\hat{t}_x)$ as for $(\hat{p}_y,\hat{t}_y)$, because of the monotonicity. This means that accuracy is of no use whatsoever in our search for the true probabilities, which correspond to $x=1$! However (drum roll), proper scoring rules like the Brier or the log score will indeed be optimized in expectation by the correct $x=1$.

brier_score <- function(probs,actuals) mean(c((1-probs)[actuals]^2,probs[!actuals]^2))
log_score <- function(probs,actuals) mean(c(-log(probs[actuals]),-log((1-probs)[!actuals])))

exponents <- 10^seq(-1,1,by=0.1)
brier_scores <- log_scores <- rep(NA,length(exponents))
for ( ii in seq_along(exponents) ) {
    brier_scores[ii] <- brier_score(true_probabilities^exponents[ii],actuals)
    log_scores[ii] <- log_score(true_probabilities^exponents[ii],actuals)
}
plot(exponents,brier_scores,log="x",type="o",xlab="Exponent",main="Brier score",ylab="")
plot(exponents,log_scores,log="x",type="o",xlab="Exponent",main="Log score",ylab="")

scores

$\endgroup$
6
  • $\begingroup$ I generally agree with the idea that if the model frequently sees a spam and a ham as separated by only a little bit, then it doesn’t really know what it’s going, but I have a bunch of follow/up questions. Let’s start with this one. How much of your objection to picking the more accurate model with the lower Brier score comes from the fact that it was only 1000 emails? What if it were on a trillion or a bazillion-gillion emails (whatever a bazillion-gillion is)? $\endgroup$
    – Dave
    Jun 3 '20 at 1:25
  • $\begingroup$ I don't see a point in my argumentation where the $n$ comes in, so I would argue in exactly the same way if $n$ were (much) larger. $\endgroup$ Jun 3 '20 at 5:56
  • $\begingroup$ I accepted this answer months ago and have pointed a number of people this way. However, I recently have become bothered by this again. Maybe spam/ham classification isn’t the right example, but for an automated task like optical character recognition, all I would care about (or would I?) is how often it gives wrong characters, perhaps considering some mistakes less important than others (e.g. O vs o). All I get from OCR is a character, not some kind of “suspected Q” flag in the subject line. I am curious about your thoughts on this example. $\endgroup$
    – Dave
    Oct 4 '20 at 15:40
  • $\begingroup$ @Dave: apologies for not getting back sooner. I guess this later thread of yours picks up where these comments left off, or is there anything left open? $\endgroup$ Nov 22 '20 at 14:33
  • $\begingroup$ I posted the other question when I realized that it would require more than a quick comment. // Ever since you posted this, I have noticed my work email giving me a "suspected spam" tag in some subject lines! $\endgroup$
    – Dave
    Nov 23 '20 at 22:22
3
$\begingroup$

I think it is worth making the distinction between performance evaluation and model selection criteria.

In performance evaluation you want to know how well your system is likely to perform in operation, based on the data you have available now. When evaluating performance, you need to use the metric that is appropriate for your task. So if you have a classification problem (i.e. you do need to make a hard "yes"/"no" decision) and the false-positive and false-negative costs are the same, then classification accuracy is an appropriate performance evaluation criterion. Obviously you would probably want to evaluate the uncertainty of the performance estimate, perhaps using bootstrap replication.

For a model selection criterion, on the other hand, where we want to chose between competing methods, accuracy might not be a good criterion, even if it is the primary quantity of interest. The problem is that accuracy is brittle, if you ran the experiment again, but with a different sample of training data, you might get a slightly different decision boundary that gave a very different accuracy on the same test data. In this case, because the margins on the first classifier are quite small, it is more likely that the decision boundary will shift if it is trained on a different sample of data than the second one. Likewise, if the classifier is evaluated on a different set of test data, the first classifier is more likely to make mistakes than the second. As the margins are small, the new test data has to be only slightly different to the old data near the decision boundary for there to be a misclassification. Thus the Brier score is rewarding the second classifier for it's greater stability. This is important as we want a classifier that will perform well in operation, not just on this particular test set.

I use least-squares support vector machine (and kernel logistic regression) a fair bit, and you need a model selection criterion for tuning the hyper-parameters. The obvious thing to do is to use cross-validation accuracy as the selection criterion for problems where accuracy is the quantity of primary interest. However it is generally better to use the cross-validated Brier score (or equivalently PRESS criterion in my case), which is less brittle. You can get better accuracy from the final model by using a proper scoring rule as the model selection criterion.

I learned this from taking part in a "Performance Prediction Challenge" hosted by a conference, where you had to provide not just the predictions, but also an estimate of how accurate (balanced accuracy in this case) your model would be on the test data. So this is based on experimentation and experience - challenges like these are a good place to find out what actually works and what doesn't as there are no straw men or "operator bias" issues in the evaluation. The paper is here FWIW (preprint here).

$\endgroup$
4
  • 1
    $\begingroup$ Thank you for posting this, even a year after I accepted the other answer. This is consistent with some of my own thoughts and even what Stephan Kolassa wrote about the perfect classifier not really knowing what it is doing. $\endgroup$
    – Dave
    Jul 29 at 17:10
  • $\begingroup$ no problem, the margin thing also relates to the support vector machine, which aims to construct a classifier with as wide a margin between the classes for pretty much the same reasons, so it also has some justification from the computational learning theory angle. $\endgroup$ Jul 29 at 17:13
  • 1
    $\begingroup$ Perhaps the answer to my boss would be along the lines of your "This is important as we want a classifier that will perform well in operation, not just on this particular test set" comment. $\endgroup$
    – Dave
    Jul 29 at 17:16
  • $\begingroup$ The bootstrap interval on the accuracy estimate from the bootstrapping would probably help as well (as long as a new classifier is trained on each replication). This also means you can make a bagged classifier, which may work a bit better anyway. $\endgroup$ Jul 29 at 17:20

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.