There are lots of applications to machine learning, and the most popular problem in practice is binary classification. Examples of things we want to predict:

  • user will click / buy something or not
  • page is appropriate to request or not
  • charge of particle is positive or negative
  • observed signal decay or something else
  • bright object on the sky is galaxy or quasar

There are many different area-specific metrics to estimate quality of classification, however the basic tool one should be able to work with regardless of the area is ROC curve (which I will talk about in this post).

Notions in binary classification for binary predictions, is/as notation

We have two classes: class 0 and class 1, background and signal respectively. In the simplest case predictions are binary: each event is attributed by classifier to be signal or background.

Unfortunately, there are too many terms used to describe this (trivial) classification result. Scary image from wikipedia shows how many different terms people were able to invent to describe a space with 4 degrees of freedom.

Let me introduce my own notion, more systematic (%picture about 15 standards%):

  • isS (isSignal), isB (isBackground) - how many events really belong to this class
  • asS (asSignal), asB (asBackground) - how many events were classified as signal (background)
  • isSasB (isSignalasBackground) - how many signal events were erroneously classified as background. isSasS, isBasB, isBasS are defined in the same way.

Hardly one can misunderstand what each of introduced numbers means.

So, there are actually only 4 basic numbers: isSasS, isSasB, isBasS, isBasB.

All the other information can be easily reconstructed:

  • isS = isSasS + isSasB
  • isB = isBasS + isBasB
  • asS = isSasS + isBasS
  • asB = isSasB + isBasB

Other typically used measures:

  • true positive rate (part of correctly classified signal, also known as recall, sensitivity or signal efficiency).

    tpr = isSasS / isS 
    
  • false positive rate (part of incorrectly classified background, aka background efficiency)

    fpr = isBasS / isB 
    
  • also there are tnr = isSasB / isS and fnr = isSasB / isS (tnr also known as specificity)

Other way to define parameters is True/False Positives/Negatives:

TP = isSasS, FP = isBasS, TN = isBasB, FN = isSasB  

But it is too easy (at least for me) to mess FP and FN, so I avoid this notion.

Continuous predictions. ROC curve

So, there are many different ways to measure quality of binary predictions. The bad thing is people frequently start by comparing them.

Meanwhile, the output of classifier is real number, not binary.

What’s wrong with using binary predictions?

  1. to estimate quality, one needs to select threshold. Either people forget about it (and use default, which is not even close to optimal one)
  2. binary metrics are frequently unstable and need many samples in validation (statistical variation is comparable to gain obtained by changing parameters)
  3. in many cases, trained classifier is not used to make decisions, but needed reconstruct the probabilities later used in next stages of ML pipeline
It's a bad idea to use rough predictions of classifier (classifier.predict(X) in scikit-learn), instead always use probabilities classifier.predict_proba(X)

So the right way is to look at the whole picture and compare how well the classifier was able to split classes. How ‘far’ distributions of output of signal and background.

ROC curve

The graphical way to compare output of two classifiers is ROC curve, which is built by checking all possible thresholds. For each threshold tpr and fpr are computed (which part of signal/background event passes this threshold).

By checking all possible thresholds, we get the ROC curve. When ROC curve coincides with diagonal — this is the worst situation, because two distributions coincide. The higher ROC curve — the better discrimination between signal and background.

If at every point ROC curve of classifier A is higher than curve of classifier B, we are sure to say that in any application classifier A is better.

ROC curve interactive demonstration interface

ROC curve interactive demo

You can play with this demonstration, it’s expected to work in any browser without additional plugins.

Instructions

On the right plot: two compared distributions (for demonstration normal distributions were taken). On the left plot: corresponding ROC curve.

Controls:

  • 4 inputs with means and variances of distributions
  • Also vertical line corresponding to threshold can be moved</li>

Presentation was prepared by Oleg Alenkin and Alex Rogozhnikov.

Area under ROC

General-purpose measure of classification quality is area under ROC curve.

In the worst case it is 0.5, while the ideal classification corresponds to area = 1.

This figure of merit is very stable (), and moreover enjoys the following mathematical property:

, where $x$ and $y$ are predictions of random signal and background samples. So this is a probability that random signal and random background samples’ predictions will be correctly ordered.

Measures of quality through the prism of ROC

Interesting moment in ROC curve is that it is completely invariant to any monotonic transformations of classifiers output. If you divide the predictions of classifier by two, ROC curve will stay the same. Or if you apply exponent — again, nothing changes.

This notable property makes ROC curve the universal base for comparison of classification models, because it contains all necessary information (passing dependencies) while ignoring everything inessential (predictions values).

However curve is not a number, and the final quality still should be computed according to real problem (based on ROC curve).

accuracy: assuming that you have $N_s$ signal objects and $N_b$ background objects, the expected number of quest samples is $N_s \times TPR + N_b \times (1 - FPR)$, to maximize the number of correctly predicted objects, you should select a threshold, which corresponds to maximum of this value on the ROC. (As you can check, at this point ROC should be tangent to line $N_s \times TPR - N_b \times FPR = const$.

weighted accuracy: another example (which frequently corresponds to practice). Imaging that for each correct guess about new object you are paid. But the cost for correctly guessing signal and background are different: $c_s, c_b$. To maximize the income, the threshold for decision rule should correspond to point at ROC curve, where is maximal.

In particle physics more complicated things are usually taken as final measure of classification model. Those are corresponding to hypothesis testing (punzy metric, median significance), but the idea is still the same: one should maximize some figure of merit and take corresponding threshold.