See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/260126882

F1-Optimal Thresholding in the Multi-Label Setting

Article · February 2014

Source: arXiv

CITATIONS READS

3 1,252

3 authors, including:

Zachary Chase Lipton Charles Elkan

Carnegie Mellon University University of California, San Diego
159 PUBLICATIONS 9,320 CITATIONS 158 PUBLICATIONS 22,124 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Deep Learning for Medical Time Series View project

AI Safety View project

All content following this page was uploaded by Zachary Chase Lipton on 01 April 2015.

The user has requested enhancement of the downloaded file.

 Thresholding Classifiers to Maximize F1 Score

Zachary C. Lipton, Charles Elkan, and Balakrishnan Naryanaswamy

University of California, San Diego,

La Jolla, California, 92093-0404, USA
{zlipton,celkan,muralib}@cs.ucsd.edu

Abstract. This paper provides new insight into maximizing F1 scores

in the context of binary classification and also in the context of multil-
arXiv:1402.1892v2 [stat.ML] 14 May 2014

abel classification. The harmonic mean of precision and recall, F1 score

is widely used to measure the success of a binary classifier when one
class is rare. Micro average, macro average, and per instance average F1
scores are used in multilabel classification. For any classifier that pro-
duces a real-valued output, we derive the relationship between the best
achievable F1 score and the decision-making threshold that achieves this
optimum. As a special case, if the classifier outputs are well-calibrated
conditional probabilities, then the optimal threshold is half the optimal
F1 score. As another special case, if the classifier is completely uninfor-
mative, then the optimal behavior is to classify all examples as positive.
Since the actual prevalence of positive examples typically is low, this
behavior can be considered undesirable. As a case study, we discuss the
results, which can be surprising, of applying this procedure when pre-
dicting 26,853 labels for Medline documents.

Keywords: machine learning, evaluation methodology, F1-score, multilabel clas-

sification, binary classification

1 Introduction

Performance metrics are useful for comparing the quality of predictions across
systems. Some commonly used metrics for binary classification are accuracy,
precision, recall, F1 score, and Jaccard index [15]. Multilabel classification is
an extension of binary classification that is currently an area of active research
in supervised machine learning [18]. Micro averaging, macro averaging, and per
instance averaging are three commonly used variants of F1 score used in the
multilabel setting. In general, macro averaging increases the impact on final
score of performance on rare labels, while per instance averaging increases the
importance of performing well on each example [17]. In this paper, we present
theoretical and experimental results on the properties of the F1 metric.1
1
For concreteness, the results of this paper are given specifically for the F1 metric
and its multilabel variants. However, the results can be generalized to Fβ metrics
for β 6= 1.
2 Zachary C. Lipton, Charles Elkan, and Balakrishnan Naryanaswamy

Two approaches exist for optimizing performance on F1. Structured loss min-
imization incorporates the performance metric into the loss function and then
optimizes during training. In contrast, plug-in rules convert the numerical out-
puts of a classifier into optimal predictions [5]. In this paper, we highlight the
latter scenario to differentiate between the beliefs of a system and the predictions
selected to optimize alternative metrics. In the multilabel case, we show that the
same beliefs can produce markedly dissimilar optimally thresholded predictions
depending upon the choice of averaging method.
That F1 is asymmetric in the positive and negative class is well-known. Given
complemented predictions and actual labels, F1 may award a different score.
It also generally known that micro F1 is affected less by performance on rare
labels, while Macro-F1 weighs the F1 of on each label equally [11]. In this pa-
per, we show how these properties are manifest in the optimal decision-making
thresholds and introduce a theorem to describe that threshold. Additionally,
we demonstrate that given an uninformative classifier, optimal thresholding to
maximize F1 predicts all instances positive regardless of the base rate.
While F1 is widely used, some of its properties are not widely recognized.
In particular, when choosing predictions to maximize the expectation of F1 for
a batch of examples, each prediction depends not only on the probability that
the label applies to that example, but also on the distribution of probabilities
for all other examples in the batch. We quantify this dependence in Theorem 1,
where we derive an expression for optimal thresholds. The dependence makes it
difficult to relate predictions that are optimally thresholded for F1 to a system’s
predicted probabilities.
We show that the difference in F1 score between perfect predictions and
optimally thresholded random guesses depends strongly on the base rate. As
a result, assuming optimal thresholding and a classifier outputting calibrated
probabilities, predictions on rare labels typically gets a score between close to
zero and one, while scores on common labels will always be high. In this sense,
macro average F1 can be argued not to weigh labels equally, but actually to give
greater weight to performance on rare labels.
As a case study, we consider tagging articles in the biomedical literature with
MeSH terms, a controlled vocabulary of 26,853 labels. These labels have hetero-
geneously distributed base rates. We show that if the predictive features for rare
labels are lost (because of feature selection or another cause) then the optimal
threshold to maximize macro F1 leads to predicting these rare labels frequently.
For the case study application, and likely for similar ones, this behavior is far
from desirable.

2 Definitions of Performance Metrics

Consider binary classification in the single or multilabel setting. Given training

data of the form {hx1 , y1 i, . . . , hxn , yn i} where each xi is a feature vector of
dimension d and each yi is a binary vector of true labels of dimension m, a
probabilistic classifier outputs a model which specifies the conditional probability
 Thresholding Classifiers to Maximize F1 Score 3

Actual Positive Actual Negative

Predicted Positive tp fp
Predicted Negative fn tn

Fig. 1: Confusion Matrix

of each label applying to each instance given the feature vector. For a batch of
data of dimension n × d, the model outputs an n × m matrix C of probabilities.
In the single-label setting, m = 1 and C is an n × 1 matrix, i.e. a column vector.
A decision rule D(C) : Rn×m → {0, 1}n×m converts a matrix of probabilities
C to binary predictions P . The gold standard G ∈ Rn×m represents the true
values of all labels for all instances in a given batch. A performance metric M
assigns a score to a prediction given a gold standard:
M (P |G) : {0, 1}n×m × {0, 1}n×m → R ∈ [0, 1].
The counts of true positives tp, false positives f p, false negatives f n, and true
negatives tn are represented via a confusion matrix (Figure 1).
Precision p = tp/(tp + f p) is the fraction of all positive predictions that are
true positives, while recall r = tp/(tp + f n) is the fraction of all actual positives
that are predicted positive. By definition the F1 score is the harmonic mean of
precision and recall: F 1 = 2/(1/r + 1/p). By substitution, F1 can be expressed
as a function of counts of true positives, false positives and false negatives:
2tp
F1 = . (1)
2tp + f p + f n
The harmonic mean expression for F1 is undefined when tp = 0, but the trans-
lated expression is defined. This difference does not impact the results below.

2.1 Basic Properties of F1

Before explaining optimal thresholding to maximize F1, we first discuss some
properties of F1. For any fixed number of actual positives in the gold standard,
only two of the four entries in the confusion matrix (Figure 1) vary independently.
This is because the number of actual positives is equal to the sum tp + f n while
the number of actual negatives is equal to the sum tn + f p. A second basic
property of F1 is that it is non-linear in its inputs. Specifically, fixing the number
f p, F1 is concave as a function of tp (Figure 2). By contrast, accuracy is a linear
function of tp and tn (Figure 3).
As mentioned in the introduction, F1 is asymmetric. By this, we mean that
the score assigned to a prediction P given gold standard G can be arbitrarily
different from the score assigned to a complementary prediction P c given com-
plementary gold standard Gc . This can be seen by comparing Figure 2 with
Figure 5. This asymmetry is problematic when both false positives and false
negatives are costly. For example, F1 has been used to evaluate the classification
of tumors as benign or malignant [1], a domain where both false positives and
false negatives have considerable costs.
4 Zachary C. Lipton, Charles Elkan, and Balakrishnan Naryanaswamy

Base Rate of 0.1 Base Rate of 0.1

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
F1 score

Accuracy
0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1 0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1
True Positive True Positive

Fig. 2: Holding base rate and f p con- Fig. 3: Unlike F1, accuracy offers lin-
stant, F1 is concave in tp. Each line early increasing returns. Each line is
is a different value of f p. a fixed value of f p.

2.2 Multilabel Performance Measures

While F1 was developed for single-label information retrieval, as mentioned there
are variants of F1 for the multilabel setting. Micro F1 treats all predictions on
all labels as one vector and then calculates the F1 score. In particular,
m
n X
1(Pij = 1)1(Gij = 1).
X
tp = 2
i=1 j=1

We define f p and f n analogously and calculate the final score using (1). Macro
F1, which can also be called per label F1, calculates the F1 for each of the m
labels and averages them:
m
1X
F 1M acro (P |G) = F 1(P:j , G:j ).
m j=1

Per instance F1 is similar but averages F1 over all n examples:

n
1X
F 1Instance (P |G) = F 1(Pi: , Gi: ).
n i=1

Accuracy is the fraction of all instances that are predicted correctly:

tp + tn
Acc = .
tp + tn + f p + f n
Accuracy is adapted to the multilabel setting by summing tp and tn for all labels
and then dividing by the total number of predictions:
n m
1 XX
Acc(P |G) = 1(Pij = Gij ).
nm i=1 j=1
 Thresholding Classifiers to Maximize F1 Score 5

Base Rate 0.1

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
0.1
0.08 1
0.06 0.8
0.04 0.6
0.4
0.02 0.2
0 0
True Positive False Positive

Fig. 4: For fixed base rate, F1 is a non-linear function with only two degrees of
freedom.

Jaccard Index, a monotonically increasing function of F1, is the ratio of the

intersection of predictions and gold standard to their union:
tp
Jaccard = .
tp + f n + f p

3 Prior Work

Motivated by the widespread use of F1 in information retrieval and in single

and multilabel binary classification, researchers have published extensively on
its optimization. [8] propose an outer-inner maximization technique for F1 max-
imization, and [4] study extensions to the multilabel setting, showing that simple
threshold search strategies are sufficient when individual probabilistic classifiers
are independent. Finally, [6] describe how the method of [8] can be extended to
efficiently label data points even when classifier outputs are dependent. More
recent work in this direction can be found in [19]. However, none of this work
directly identifies the relationship of optimal thresholds to the maximum achiev-
able F1 score over all thresholds, as we do here.
While there has been work on applying general constrained optimization
techniques to related metrics [13], research often focuses on specific classification
methods. In particular, [16] study F1 optimization for conditional random fields
and [14] perform the same optimization for SVMs. In our work, we study the
consequences of such optimization for probabilistic classifiers, particularly in the
multilabel setting.
A result similar to our special case (Corollary 1) was recently derived in
[20]. However, their derivation is complex and does not prove our more general
6 Zachary C. Lipton, Charles Elkan, and Balakrishnan Naryanaswamy

Base Rate of 0.1

1

0.9
1
0.8
0.9
0.7
0.8

0.6 0.7
F1 score

Expected F1 Score
0.5 0.6

0.4 0.5

0.3 0.4

0.3
0.2
0.2
0.1
0.1
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0
True Negative 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Percent Predicted Positive

Fig. 5: F1 score for fixed base rate

Fig. 6: The expected F1 score of an
and number f n of false negatives. F1
optimally thresholded random guess
offers increasing marginal returns as
is highly dependent on the base rate.
a function of tn. Each line is a fixed
value of f n.

Theorem 1 which describes the optimal decision-making threshold even when

the scores output by a classifier are not probabilities. Their paper also does not
contain the empirical version we derive for the multilabel setting in Theorem 2.
The batch observation is related to the observation in [9] that given some
classifier, a specific example may or may not cross the decision threshold, de-
pending on the other examples present in the test data. However, they do not
identify this threshold as F21 or make use of this fact to explain the differences
between predictions made to optimize micro and macro average F1.

4 Optimal Decision Regions for F1 Maximization

In this section, we provide a characterization of the optimal decision regions
that maximize F1 and, for a special case, we present a relationship between the
optimal threshold and the maximum achievable F1 score.
We assume that the classifier outputs real-valued scores s and that there exist
two distributions p(s|t = 1) and p(s|t = 0) that are the conditional probability
of seeing the score s when the true label t is 1 or 0, respectively. We assume
that these distributions are known in this section; the next section discusses an
empirical version of the result. Note also that in this section tp etc. are fractions
that sum to one, not counts.
Given p(s|t = 1) and p(s|t = 0), we seek a decision rule D : s → {0, 1}
mapping scores to class labels such that the resultant classifier maximizes F1.
We start with a lemma that is valid for any D.
R
Lemma 1. The true positive rate tp = b s:D(s)=1 p(s|t = 1)ds where b = p(t =
1) is the base rate.
 Thresholding Classifiers to Maximize F1 Score 7
R
Proof. Clearly tp = s:D(s)=1 p(t = 1|s)p(s)ds. Bayes rule says that p(t = 1|s) =
R
p(s|t = 1)p(t = 1)/p(s). Hence tp = b s:D(s)=1 p(s|t = 1)ds.
Using three similar lemmas, the entries of the confusion matrix are
Z
tp = b p(s|t = 1)ds
s:D(s)=1
Z
fn = b p(s|t = 1)ds
s:D(s)=0

Z
f p = (1 − b) p(s|t = 0)ds
s:D(s)=1
Z
tn = (1 − b) p(s|t = 0)ds.
s:D(s)=0

The following theorem describes the optimal decision rule that maximizes F1.
Theorem 1. A score s is assigned to the positive class, that is D(s) = 1, by a
classifier that maximizes F1 if and only if
b · p(s|t = 1)
≥J (2)
(1 − b) · p(s|t = 0)
tp
where J = f n+tp+f p is the Jaccard index of the optimal classifier, with ambiguity
given equality in (2).
Before we provide the proof of this theorem, we note the difference between
the rule in (2) and conventional cost-sensitive decision making [7] or Neyman-
Pearson detection. In the latter, the right hand side J is replaced by a constant
λ that depends only on the costs of 0 − 1 and 1 − 0 classification errors, and not
on the performance of the classifier on the entire batch. We will later elaborate
on this point, and describe how this relationship leads to potentially undesirable
thresholding behavior for many applications in the multilabel setting.
Proof. Divide the domain of s into regions of size ∆. Suppose that the decision
rule D(·) has been fixed for all regions except a particular region R denoted ∆
around a point (with some abuse of notation) s. Write P1 (∆) = ∆ p(s|t = 1)
and define P0 (∆) similarly.
Suppose that the F1 achieved with decision rule D for all scores besides D(∆)
2tp
is F 1 = 2tp+f n+f p . Now, if we add ∆ to the positive part of the decision rule,
D(∆) = 1, then the new F1 score will be
2tp + 2bP1 (∆)
F 10 = .
2tp + 2bP1 (∆) + f n + f p + (1 − b)P0 (∆)
On the other hand, if we add ∆ to the negative part of the decision rule, D(∆) =
0, then the new F1 score will be
2tp
F 100 = .
2tp + f n + bP1 (∆) + f p
8 Zachary C. Lipton, Charles Elkan, and Balakrishnan Naryanaswamy

We add ∆ to the positive class only if F 10 ≥ F 100 . With some algebraic simpli-
fication, this condition becomes

bP1 (∆) tp
≥ .
(1 − b)P0 (∆) tp + f n + f p

Taking the limit |∆| → 0 gives the claimed result.

If, as a special case, the model outputs calibrated probabilities, that is p(t =
1|s) = s and p(t = 0|s) = 1 − s, then we have the following corollary.

Corollary 1. An instance with predicted probability s is assigned to the positive

class by the optimal decision rule that maximizes F1 if and only if s ≥ F/2 where
2tp
F = 2tp+f n+f p is the F1 score achieved by this optimal decision rule.

Proof. Using the definition of calibration and then Bayes rule, for the optimal
decision surface for assigning a score s to the positive class

p(t = 1|s) s p(s|t = 1)b

= = . (3)
p(t = 0|s) 1−s p(s|t = 0)(1 − b)

Incorporating (3) in Theorem 1 gives

s tp
≥ .
1−s f n + tp + f p

Simplifying results in
tp F
s≥ = .
2tp + f n + f p 2
Thus, the optimal threshold in the calibrated case is half the maximum F 1.
Above, we assume that scores have a distribution conditioned on the true
class. Using the intuition in the proof of Theorem 1, we can also derive an
empirical version of the result. To save space, we provide a more general version
of the empirical result in the next section for multilabel problems, noting that a
similar non-probabilistic statement holds for the single label setting as well.

4.1 Maximizing Expected F1 Using a Probabilistic Classifier

The above result can be extended to the multilabel setting with dependence. We
give a different proof that confirms the optimal threshold for empirical maxi-
mization of F1.
We first present an algorithm from [6]. Let s be the output vector of length n
scores from a model, to predict n labels in the multilabel setting. Let t ∈ {0, 1}n
be the gold standard and h ∈ {0, 1}n be the thresholded output for a given set
of n labels. In addition, define a = tp + f n, the total count of positive labels in
the gold standard and c = tp + f p the total count of predicted positive labels.
Note that a and c are functions of t and h, though we suppress this dependence
 Thresholding Classifiers to Maximize F1 Score 9

in notation. Define za =
P
t:tp+f n=a tp(t). The maximum achievable macro F1
is
 
2tp
F 1 = max max Ep(t|s)
c h:tp+f p=c 2tp + f p + f n
X za
= max max 2hT .
c h:tp+f p=c
a
a+c

P Algorithm:
za
Loop over the number of predicted positives c. Sort the vector
a a+c of length n. Proceed along its entries one by one. Adding an entry to
the positive class increases the numerator by za , which is always positive. Stop
after entry number c. Pick the c value and corresponding threshold which give
the largest F1.
Some algebra gives the following interpretation:
X E(tp|c)
max E(F 1) = max p(a).
c c
a
a+c

Theorem 2. The stopping threshold will be max Ep(y|s) [ F21 ].

4.2 Consequences of F1 Optimal Classifier Design

We demonstrate two consequences of designing classifiers that maximize F1.

These are the “batch observation” and the “uninformative classifier observation.”
We will later demonstrate with a case study that these can combine to produce
surprising and potentially undesirable optimal predictions when macro F1 is
optimized in practice.
The batch observation is that a label may or may not be predicted for an
instance depending on the distribution of other probabilities in the batch. Earlier,
we observed a relationship between the optimal threshold and the maximum
E(F 1) and demonstrated that the maximum E(F 1) is related to the distribution
of probabilities for all predictions. Therefore, depending upon the distribution
in which an instance is placed, it may or may not exceed the optimal threshold.
Note that because F1 can never exceed 1, the optimal threshold can never exceed
.5.
Consider for example an instance with probability 0.1. It will be predicted
positive if it has the highest probability of all instances in a batch. However, in
a different batch, where the probabilities assigned to all other elements are 0.5
and n is large, the maximum E(F 1) would be close to 2/3. According to the
theorem, we will predict positive on this last instance only if it has a probability
greater than 1/3.
An uninformative classifier is one that assigns the same score to all examples.
If these scores are calibrated probabilities, the base rate is assigned to every
example.
10 Zachary C. Lipton, Charles Elkan, and Balakrishnan Naryanaswamy

Theorem 3. Given an uninformative classifier for a label, optimal thresholding

to maximize F1 results in predicting all examples positive.

Proof. Given an uninformative classifier, we seek the optimal threshold that

maximizes E(F 1). The only choice is how many labels to predict. By symmetry
between the instances, it doesn’t matter which instances are labeled positive.
Let a = tp + f n be the number of actual positives and let c = tp + f p be
the number of positive predictions. The denominator of the expression for F1 in
Equation (1), that is 2tp + f p + f n = a + c, is constant. The number of true
positives, however, is a random variable. Its expected value is equal to the sum
of the probabilities that each example predicted positive actually is positive:
Pc
2 i=1 b 2c · b
E(F 1) = =
a+c a+c
where b = a/n is the base rate. To maximize this expectation as a function of c,
we calculate the partial derivative with respect to c, applying the product rule:

∂ ∂ 2c · b 2b 2c · b
E(F 1) = = − .
∂c ∂c a + c a + c (a + c)2

Both terms in the difference are always positive, so we can show that this deriva-
tive is always positive by showing that

2b 2c · b
> .
a+c (a + c)2
c
Simplification gives the condition 1 > a+c . As this condition always holds, the
derivative is always positive. Therefore, whenever the frequency of actual posi-
tives in the test set is nonzero, and the classifier is uninformative, expected F1
is maximized by predicting that all examples are positive.

For low base rates an optimally thresholded uninformative classifier achieves

E(F 1) close to 0, while for high base rates E(F 1) is close to 1 (Figure 6). We
revisit this point in the context of macro F1.

5 Multilabel Setting

Different metrics are used to measure different aspects of a system’s performance.

However, by changing the loss function, this can change the optimal predictions.
We relate the batch observation to discrepancies between predictions optimal for
micro and macro F1. We show that while micro F1 is dominated by performance
on common labels, macro F1 disproportionately weights rare labels. Addition-
ally, we show that macro averaging over F1 can conceal uninformative classifier
thresholding.
Consider the equation for F1, and imagine tp, f p, and f n to be known for
m − 1 labels with some distribution of base rates. Now consider the mth label
 Thresholding Classifiers to Maximize F1 Score 11

to be rare with respect to the distribution. A perfect classifier increases tp by a

small amount ε equal to the number b · n of actual positives for that rare label,
while contributing nothing to the counts f p or f n:
2(tp + b · n)
F 10 = .
2(tp + b · n) + f p + f n
On the other hand, a trivial prediction of all negative only increases f n by a
small amount:
2tp
F 100 = .
2tp + f p + (f n + b · n)
By contrast, predicting all positive for a rare label will increase f p by a large
amount β = n − ε. We have

F 10 1 + b·n
tp
= .
F 100 nb
1 + a+c+b·n

where a and c are the number of positives in the gold standard and the Pnumber
of positive Ppredictions for the first m − 1 labels. We have a + c ≤ n i bi and
so if bm  i bi this ratio is small. Thus, performance on rare labels is washed
out.
In the single-label setting, the small range between the F1 value achieved by a
trivial classifier and a perfect one may not be problematic. If a trivial system gets
a score of 0.9, we can adjust the scale for what constitutes a good score. However,
when averaging separately calculated F1 over all labels, this variability can skew
scores to disproportionately weight performance on rare labels. Consider the two
label case when one label has a base rate of 0.5 and the other has a base rate
of 0.1. The corresponding expected F1 for trivial classifiers are 0.67 and 0.18
respectively. Thus the expected F1 for optimally thresholded trivial classifiers is
0.42. However, an improvement to perfect predictions on the rare label elevates
the macro F1 to 0.84 while such an improvement on the common label would
only correspond to a macro F1 of 0.59. Thus the increased variability of F1
results in high weight for rare labels in macro F1.
For a rare label with an uninformative classifier, micro F1 is optimized by
predicting all negative while macro is optimized by predicting all positive. Ear-
lier, we proved that the optimal threshold for predictions based on a calibrated
probabilistic classifier is half of the maximum F1 attainable given any thresh-
old setting. In other words, which batch an example is submitted with affects
whether a positive prediction will be made. In practice, a system may be tasked
with predicting labels with widely varying base rates. Additionally a classifier’s
ability to make confident predictions may vary widely from label to label.
Optimizing micro F1 as compared to macro F1 can be thought of as choosing
optimal thresholds given very different batches. If the base rate and distribution
of probabilities assigned to instances vary from label to label, so will the predic-
tions. Generally, labels with low base rates and less informative classifiers will
be over-predicted to maximize macro F1 as compared to micro F1. We present
empirical evidence of this phenomenon in the following case study.
12 Zachary C. Lipton, Charles Elkan, and Balakrishnan Naryanaswamy

MeSH Term Count Max F1 Threshold

Humans 2346 0.9160 0.458
Male 1472 0.8055 0.403
Female 1439 0.8131 0.407
Phosphinic Acids 1401 1.544 · 10−4 7.71 · 10−5
Penicillanic Acid 1064 8.534 · 10−4 4.27 · 10−4
Adult 1063 0.7004 0.350
Middle Aged 1028 0.7513 0.376
Platypus 980 4.676 · 10−4 2.34 · 10−4

Fig. 7: Frequently predicted MeSH Terms. When macro F1 is optimized, low

thresholds are set for rare labels (bold) with uninformative classifiers.

6 Case Study
This section discusses a case study that demonstrates how in practice, threshold-
ing to maximize macro-F1 can produce undesirable predictions. To our knowl-
edge, a similar real-world case of pathological behavior has not been previously
described in the literature, even though macro averaging F1 is a common ap-
proach.
We consider the task of assigning tags from a controlled vocabulary of 26,853
MeSH terms to articles in the biomedical literature using only titles and ab-
stracts. We represent each abstract as a sparse bag-of-words vector over a vo-
cabulary of 188,923 words. The training data consists of a matrix A with n rows
and d columns, where n is the number of abstracts and d is the number of fea-
tures in the bag of words representation. We apply a tf-idf text preprocessing
step to the bag of words representation to account for word burstiness [10] and
to elevate the impact of rare words.
Because linear regression models can be trained for multiple labels efficiently,
we choose linear regression as a model. Note that square loss is a proper loss
function and does yield calibrated probabilistic predictions [12]. Further, to in-
crease the speed of training and prevent overfitting, we approximate the training
matrix A by a rank restricted Ak using singular value decomposition. One po-
tential consequence of this rank restriction is that the signal of extremely rare
words can be lost. This can be problematic when rare terms are the only features
of predictive value for a label.
Given the probabilistic output of the classifier and the theory relating opti-
mal thresholds to maximum attainable F1, we designed three different plug-in
rules to maximize micro, macro and per instance F1. Inspection of the predic-
tions to maximize micro F1 revealed no irregularities. However, inspecting the
predictions thresholded to maximize performance on macro F1 showed that sev-
eral terms with very low base rates were predicted for more than a third of all
test documents. Among these terms were “Platypus”, “Penicillanic Acids” and
“Phosphinic Acids” (Figure 7).
In multilabel classification, a label can have low base rate and an uninforma-
tive classifier. In this case, optimal thresholding requires the system to predict
 Thresholding Classifiers to Maximize F1 Score 13

all examples positive for this label. In the single-label case, such a system would
achieve a low F1 and not be used. But in the macro averaging multilabel case,
the extreme thresholding behavior can take place on a subset of labels, while the
system manages to perform well overall.

7 A Winner’s Curse

In practice, decision regions that maximize F1 are often set experimentally,

rather than analytically. That is, given a set of training examples, their scores
and ground truth decision regions for scores that map to different labels are set
that maximize F1 on the training batch.
In such situations, the optimal threshold can be subject to a winner’s curse
[2] where a sub-optimal threshold is chosen because of sampling effects or lim-
ited training data. As a result, the future performance of a classifier using this
threshold is less than the empirical performance. We show that threshold opti-
mization for F1 is particularly susceptible to this phenomenon (which is a type
of overfitting).
In particular, different thresholds have different rates of convergence of es-
timated F1 with number of samples n. As a result, for a given n, comparing
the empirical performance of low and high thresholds can result in suboptimal
performance. This is because, for a fixed number of samples, some thresholds
converge to their true error rates while others have higher variance and may be
set erroneously. We demonstrate these ideas for a scenario with an uninformative
model, though they hold more generally.
Consider an uninformative model, for a label with base rate b. The model
is uninformative in the sense that output scores are si = b + ni ∀ i, where
ni = N (0, σ 2 ). Thus, scores are uncorrelated with and independent of the true
labels. The empirical accuracy for a threshold t is
1X 1X
Atexp = 1[Si ≥ t] + 1[Si ≤ t] (4)
n i∈+ n i∈−

where + and − index the positive and negative class respectively. Each term
in Equation (4) is the sum of O(n) i.i.d random variables and has exponential
(in n) rate of convergence to the mean irrespective of the base rate b and the
threshold t. Thus, for a fixed number T of threshold choices, the probability of
choosing the wrong threshold Perr ≤ T 2−n where  depends on the distance
between the optimal and next nearest threshold. Even if errors occur the most
likely errors are thresholds close to the true optimal threshold (a consequence of
Sanov’s Theorem [3]).
Consider how F1-maximizing thresholds would be set experimentally, given
a training batch of independent ground truth and scores from an uninformative
classifier. The scores si can be sorted in decreasing order (w.l.o.g.) since they are
independent of the true labels for an uninformative classifier. Based on these, we
empirically select the threshold that maximizes F1 on the training batch. The
14 Zachary C. Lipton, Charles Elkan, and Balakrishnan Naryanaswamy

optimal empirical threshold will lie between two scores that include the value
F1
2 , when the scores are calibrated, in accordance with Theorem 1.
The threshold smin that classifies all examples positive (and maximizes F 1
analytically by Theorem 3) has an empirical F1 close to its expectation of
2b 2
1+b = 1+1/b since tp, f p and f n are all estimated from the entire data. Consider
the threshold smax that classifies only the first example positive and all others
negative. With probability b, this has F1 score 2/(2 + b · n), which is lower than
that of the optimal threshold only when
q
1 + n8 − 1
b≥ .
2
Despite the threshold smax being far from optimal, it has a constant probability
of having a higher F1 on training data, a probability that does not decrease
with n, for n < (1 − b)/b2 . Therefore, optimizing F1 will have a sharp threshold
behavior, where for n < (1 − b)/b2 the algorithm will identify large thresholds
with constant probability, whereas for larger n it will correctly identify small
thresholds. Note that identifying optimal thresholds for F 1 is still problematic
since it then leads to issue identified in the previous section. While these issues
are distinct, they both arise from the nonlinearity of F1 score and its asymmetric
treatment of positive and negative labels.
We simulate this behavior, executing 10,000 runs for each setting of the base
rate, with n = 106 samples for each run to set the threshold (Figure 8). Scores
are chosen using variance σ 2 = 1. True labels are assigned at the base rate,
independent of the scores. The threshold that maximizes F1 on the training set
is selected. We plot a histogram of the fraction predicted positive as a function
of the empirically chosen threshold. There is a shift from predicting almost all
positives to almost all negatives as base rate is decreased. In particular, for low
base rate b, even with a large number of samples, a small fraction of examples
are predicted positive. The analytically derived optimal decision in all cases is
to predict all positive, i.e. to use a threshold of 0.

8 Discussion

In this paper, we present theoretical and empirical results describing the prop-
erties of the F1 performance metric for multilabel classification. We relate the
best achievable F1 score to the optimal decision-making threshold and show that
when a classifier is uninformative, predicting all instances positive maximizes the
expectation of F1. Further, we show that in the multilabel setting, this behavior
can be problematic when the metric to maximize is macro F1 and for a subset of
rare labels the classifier is uninformative. In contrast, we demonstrate that given
the same scenario, expected micro F1 is maximized by predicting all examples
to be negative. This knowledge can be useful as such scenarios are likely to occur
in settings with a large number of labels. We also demonstrate that micro F1 has
the potentially undesirable property of washing out performance on rare labels.
 Thresholding Classifiers to Maximize F1 Score 15

Base rate 0.5

1
0.1
0.05
0.01
0.8
0.001
0.0001
Fraction of runs

0.00001
0.6
0.000001

0.4

0.2

0 0.2 0.4 0.6 0.8 1

Percentage declared positive

Fig. 8: The distribution of experimentally chosen thresholds changes with varying

b. For small b, a small fraction of examples are predicted positive even though
the optimal thresholding is to predict all positive.

No single performance metric can capture every desirable property. For ex-
ample, separately reporting precision and recall is more informative than re-
porting F1 alone. Sometimes, however, it is practically necessary to define a
single performance metric to optimize. Evaluating competing systems and ob-
jectively choosing a winner presents such a scenario. In these cases, a change of
performance metric can have the consequence of altering optimal thresholding
behavior.

References
1. Akay, M.F.: Support vector machines combined with feature selection for breast
cancer diagnosis. Expert Systems with Applications 36(2), 3240–3247 (2009)
2. Capen, E.C., Clapp, R.V., Campbell, W.M.: Competitive bidding in high-risk sit-
uations. Journal of Petroleum Technology 23(6), 641–653 (1971)
3. Cover, T.M., Thomas, J.A.: Elements of information theory. John Wiley & Sons
(2012)
4. del Coz, J.J., Diez, J., Bahamonde, A.: Learning nondeterministic classifiers. Jour-
nal of Machine Learning Research 10, 2273–2293 (2009)
5. Dembczynski, K., Kotlowski, W., Jachnik, A., Waegeman, W., Hüllermeier, E.:
Optimizing the F-measure in multi-label classification: Plug-in rule approach versus
structured loss minimization. In: ICML (2013)
 16 Zachary C. Lipton, Charles Elkan, and Balakrishnan Naryanaswamy

6. Dembczyński, K., Waegeman, W., Cheng, W., Hüllermeier, E.: An exact algorithm
for F-measure maximization. In: Neural Information Processing Systems (2011)
7. Elkan, C.: The foundations of cost-sensitive learning. In: International joint con-
ference on artificial intelligence. pp. 973–978 (2001)
8. Jansche, M.: A maximum expected utility framework for binary sequence labeling.
In: Annual Meeting of the Association For Computational Linguistics. p. 736 (2007)
9. Lewis, D.D.: Evaluating and optimizing autonomous text classification systems. In:
Proceedings of the 18th annual international ACM SIGIR conference on research
and development in information retrieval. pp. 246–254. ACM (1995)
10. Madsen, R., Kauchak, D., Elkan, C.: Modeling word burstiness using the Dirichlet
distribution. In: Proceedings of the International Conference on Machine Learning
(ICML). pp. 545–552 (Aug 2005)
11. Manning, C., Raghavan, P., Schütze, H.: Introduction to information retrieval,
vol. 1. Cambridge University Press (2008)
12. Menon, A., Jiang, X., Vembu, S., Elkan, C., Ohno-Machado, L.: Predicting accurate
probabilities with a ranking loss. In: Proceedings of the International Conference
on Machine Learning (ICML) (Jun 2012)
13. Mozer, M.C., Dodier, R.H., Colagrosso, M.D., Guerra-Salcedo, C., Wolniewicz,
R.H.: Prodding the ROC curve: Constrained optimization of classifier performance.
In: NIPS. pp. 1409–1415 (2001)
14. Musicant, D.R., Kumar, V., Ozgur, A., et al.: Optimizing F-measure with support
vector machines. In: FLAIRS Conference. pp. 356–360 (2003)
15. Sokolova, M., Lapalme, G.: A systematic analysis of performance measures for
classification tasks. Information Processing and Management 45, 427–437 (2009)
16. Suzuki, J., McDermott, E., Isozaki, H.: Training conditional random fields with
multivariate evaluation measures. In: Proceedings of the 21st International Con-
ference on Computational Linguistics and the 44th annual meeting of the Associ-
ation for Computational Linguistics. pp. 217–224. Association for Computational
Linguistics (2006)
17. Tan, S.: Neighbor-weighted k-nearest neighbor for unbalanced text corpus. Expert
Systems with Applications 28, 667–671 (2005)
18. Tsoumakas, Grigorios & Katakis, I.: Multi-label classification: An overview. Inter-
national Journal of Data Warehousing and Mining 3(3), 1–13 (2007)
19. Ye, N., Chai, K.M., Lee, W.S., Chieu, H.L.: Optimizing F-measures: A tale of two
approaches. In: Proceedings of the International Conference on Machine Learning
(2012)
20. Zhao, M.J., Edakunni, N., Pocock, A., Brown, G.: Beyond Fano’s inequality:
Bounds on the optimal F-score, BER, and cost-sensitive risk and their implica-
tions. Journal of Machine Learning Research 14(1), 1033–1090 (2013)

View publication stats