Many bioinformatics algorithms can be understood as binary classifiers, as they are used to investigate whether a query entity belongs to a certain class ¹ . Score-based binary classifiers assign a number to the query. If this score surpasses a threshold, the query is assigned to the class under consideration. A minority of users are able to choose a threshold using their understanding of the algorithm, while the majority uses the default threshold.

Binary classifiers are often trained and compared under a unified framework, the receiver operating characteristic ( ROC) curve ² . Briefly, classifier output is first compared to a training set at all possible classification thresholds, yielding the confusion matrix with the number of true positives ( TP), false positives ( FP), true negatives ( TN) and false negatives ( FN) ( Table 1). The ROC curve plots the true positive rate ( TPR = TP/( TP + FN)), also called sensitivity,) against the false positive rate ( FPR = FP/( FP + TN)) , which equals 1-specificity) ( Figure 1, continuous line). Classifier training often aims at maximizing the area under the ROC curve, which amounts to maximizing the probability that a randomly chosen positive is ranked before a randomly chosen negative ² . This summary statistic measures performance without committing to a threshold.

Practical application of a classifier requires using a threshold-dependent performance measure to choose the operating point ^{1, 3} . This is in practice a complex task because the application domain may be skewed in two ways ⁴ . First, for many relevant bioinformatics problems the prevalence of positives in nature q _P = ( TP + FN)/( TP + TN + FP + FN) does not necessarily match the training set q _P and is hard to estimate ^{2, 5} . Second, the yields (or costs) for correct and incorrect classification of positives and negatives in the machine learning paradigm ( Y _TP , Y _TN , Y _FP , Y _FN ) may be different from each other and highly context-dependent ^{1, 3} . Points in the ROC plane with equal performance are connected by iso-yield lines with a slope, the skew ratio, which is the product of the class skew and the yield skew ⁴ : S K E W   R A T I O = q N . ( Y F P + Y T N ) q P . ( Y T P + Y F N ) ( 1 )

The skew ratio expresses the relative importance of negatives and positives, regardless of the source of the skew ⁴ . Multiple threshold-dependent performance measures have been proposed and discussed in terms of skew sensitivity ^{3, 4} , but often not justified from first principles.

Game theory allows us to consider a binary classifier as a zero-sum game between nature and the classifier ⁶ . In this game, nature is a player that uses a mixed strategy, with probabilities q _P and q _N =1- q _P for positives and negatives, respectively. The algorithm is the second player, and each threshold value corresponds to a mixed strategy with probabilities p _P and p _N for positives and negatives. Two of the four outcomes of the game, TP and TN, favor the classifier, while the remaining two, FP and FN, favor nature. The game payoff matrix ( Table 2) displays the four possible outcomes and the corresponding classifier utilities a, b, c and d. The Utility of the classifier within the game is: U T I L I T Y = a . T P + d . T N + b . F P + c . F N T P + T N + F P + F N ( 2 )

The payoff matrix for this zero-sum game corresponds directly to the confusion matrix for the classifier, and the game utilities a, b, c, d correspond to the machine learning yields Y _TP , Y _FP , Y _FN , Y _TN , respectively ( Table 1). Without loss of generality ⁴ , we can study the case a=d=1 and b=c=0. Classifier Utility within the game then reduces to the Accuracy or fraction of correct predictions ^{2– 4} . In sum, maximizing the Utility of a binary classifier in a zero-sum game against nature is equivalent to maximizing its Accuracy, a common threshold-dependent performance measure.

We can now use the minimax principle from game theory ⁶ to choose the operating point for the classifier. This principle maximizes utility for a player within a game using a pessimistic approach. For each possible action a player can take, we calculate a worst-case utility by assuming that the other player will take the action that gives them the highest utility (and the player of interest the lowest). The player of interest should take the action that maximizes this minimal, worst-case utility. Thus, the minimax utility of a player is the largest value that the player can be sure to get regardless of the actions of the other player.

In our classifier versus nature game, Utility/Accuracy of the classifier is skew-sensitive, depending on q _P for a given threshold ^{3, 4} : U T I L I T Y = 1 − F P R + q P . ( F P R + T P R − 1 ) ( 3 )

The derivative of the Utility with respect to q _P is zero along the TPR = 1 − FPR line in ROC space ( Figure 1, dashed line). The derivative is negative below this line and positive above it, indicating that points along this line are minima of the Utility function with respect to the strategy q _P of the nature player. According to the minimax principle, the classifier player should operate at the point along the TPR = 1 − FPR line that maximizes Utility. In ROC space, this condition corresponds to the intersection between the ROC curve and the descending diagonal ( Figure 1, empty circle) and yields a minimax value of 1 − FPR for the Utility. It is worth noting that this analysis regarding class skew is also valid for yield/cost skew ⁴ .

We showed that binary classifiers may be analyzed in terms of game theory. From the minimax principle, we propose a criterion to choose an operating point for the classifier that maximizes robustness against uncertainties in the skew ratio, i.e., in the prevalence of positives in nature and in yield skew, i.e., the yields/costs for true positives, true negatives, false positives and false negatives. This can be of practical value, since these uncertainties are widespread in bioinformatics and clinical applications.

In machine learning theory, TPR = 1 − FPR is the line of skew-indiference for Accuracy as a performance metric ⁴ . This is in agreement with the skew-indifference condition imposed by the minimax principle from game theory. However, to our knowledge, skew-indifference has not been exploited for optimal threshold estimation. Furthermore, the operating point of a classifier is often chosen by balancing sensitivity and specificity, without reference to the rationale behind ⁷ . Our game theory analysis shows that this empirical practice can be understood as a maximization of classifier robustness.