Surrogate regret bounds for generalized classification performance metrics
Abstract
We consider optimization of generalized performance metrics for binary classification by means of surrogate losses. We focus on a class of metrics, which are linear-fractional functions of the false positive and false negative rates (examples of which include \(F_{\beta }\)-measure, Jaccard similarity coefficient, AM measure, and many others). Our analysis concerns the following two-step procedure. First, a real-valued function f is learned by minimizing a surrogate loss for binary classification on the training sample. It is assumed that the surrogate loss is a strongly proper composite loss function (examples of which include logistic loss, squared-error loss, exponential loss, etc.). Then, given f, a threshold \(\widehat{\theta }\) is tuned on a separate validation sample, by direct optimization of the target performance metric. We show that the regret of the resulting classifier (obtained from thresholding f on \(\widehat{\theta }\)) measured with respect to the target metric is upperbounded by the regret of f measured with respect to the surrogate loss. We also extend our results to cover multilabel classification and provide regret bounds for micro- and macro-averaging measures. Our findings are further analyzed in a computational study on both synthetic and real data sets.
Keywords
Generalized performance metric Regret bound Surrogate loss function Binary classification Multilabel classification F-measure Jaccard similarity AM measure1 Introduction
In binary classification, misclassification error is not necessarily an adequate evaluation metric, and one often resorts to more complex metrics, better suited for the problem. For instance, when classes are imbalanced, \(F_{\beta }\)-measure (Lewis 1995; Jansche 2005; Nan et al. 2012) and AM measure (balanced error rate) (Menon et al. 2013) are frequently used. Optimizing such generalized performance metrics poses computational and statistical challenges, as they cannot be decomposed into losses on individual observations.
In this paper, we consider optimization of generalized performance metrics by means of surrogate losses. We restrict our attention to a family of performance metrics which are ratios of linear functions of false positives (FP) and false negatives (FN). Such functions are called linear-fractional, and include the aforementioned \(F_{\beta }\) and AM measures, as well as Jaccard similarity coefficient, weighted accuracy, and many others (Koyejo et al. 2014, 2015). We focus on the most popular approach to optimizing generalized performance metrics in practice, based on the following two-step procedure. First, a real-valued function f is learned by minimizing a surrogate loss for binary classification on the training sample. Then, given f, a threshold \(\widehat{\theta }\) is tuned on a separate validation sample, by direct optimization of the target performance measure with respect to a classifier obtained from f by thresholding at \(\widehat{\theta }\), classifying all observations with value of f above the threshold as positive class, and all observations below the threshold as negative class. This approach can be motivated by the asymptotic analysis: minimization of appropriate surrogate loss results in estimation of conditional (“posterior”) class probabilities, and many performance metrics are maximized by a classifier which predicts by thresholding on the scale of conditional probabilities (Nan et al. 2012; Zhao et al. 2013; Koyejo et al. 2014). However, it is unclear what can be said about the behavior of this procedure on finite samples.
In this work, we are interested in theoretical analysis and justification of this approach for any sample size, and for any, not necessarily perfect, classification function. To this end, we use the notion of regret with respect to some evaluation metric, which is the difference between the performance of a given classifier and the performance of the optimal classifier with respect to this metric. We show that the regret of the resulting classifier (obtained from thresholding f on \(\widehat{\theta }\)) measured with respect to the target metric is upperbounded by the regret of f measured with respect to the surrogate loss. Our result holds for any surrogate loss function, which is strongly proper composite (Agarwal 2014), examples of which include logistic loss, squared-error loss, exponential loss, etc. Interestingly, the proof of our result goes by an intermediate bound of the regret with respect to the target measure by a cost-sensitive classification regret. As a byproduct, we get a bound on the cost-sensitive classification regret by a surrogate regret of a real-valued function which holds simultaneously for all misclassification costs: the misclassification costs only influence the threshold, but not: the function, the surrogate loss, or the regret bound.
We further extend our results to cover multilabel classification, in which the goal is to simultaneously predict multiple labels for each object. We consider two methods of generalizing binary classification performance metrics to the multilabel setting: the macro-averaging and the micro-averaging (Manning et al. 2008; Parambath et al. 2014; Koyejo et al. 2015). The macro-averaging is based on first computing the performance metric separately for each label, and then averaging the metrics over the labels. In the micro-averaging, the false positives and false negatives for each label are first averaged over the labels, and then the performance metric is calculated on these averaged quantities. We show that our regret bounds hold for both macro- and micro-averaging measures. Interestingly, for micro averaging, only a single threshold needs to be tuned and is shared among all labels.
Our finding is further analyzed in a computational study on both synthetic and real data sets. We compare the performance of the algorithm when used with two types of surrogate losses: the logistic loss (which is strongly proper) and the hinge loss (which is not a proper loss). On synthetic data sets, we analyze the behavior of the algorithm for discrete feature distribution (where nonparametric classifiers are used), and for continuous feature distribution (where linear classifiers are used). Next, we look at the performance of the algorithm on real-life benchmark data sets, both for binary and multilabel classification.
We note that the goal of this paper is not to propose a new learning algorithm, but rather to provide a deeper statistical understanding of an existing method. The two-step procedure analyzed here (also known as the plug-in method in the case when the outcomes of the function have a probabilistic interpretation) is commonly used in binary classification under generalized performance metrics, but this is exactly the reason why we think it is important to study this method in more depth from a theoretical point of view.
1.1 Related work
In machine learning, numerous attempts to optimize generalized performance metrics have been proposed. They can be divided into two general categories. The structured loss approaches (Musicant et al. 2003; Tsochantaridis et al. 2005; Petterson and Caetano 2011, 2010) rely on incorporating the performance metric into the training process, thus requiring specialized learning algorithms to optimize non-standard objectives. On the other hand, the plug-in approaches, which are very closely related to the topic of this work, are based on obtaining reliable class conditional probability estimates by employing standard algorithms minimizing some surrogate loss for binary classification (such as logistic loss used in logistic regression, exponential loss used in boosting, etc.), and then plugging these estimates into the functional form of the optimal prediction rule for a given performance metric (Jansche 2007; Nan et al. 2012; Dembczyński et al. 2013; Waegeman et al. 2013; Narasimhan et al. 2014, 2015; Koyejo et al. 2014, 2015).
Existing theoretical work on generalized performance metrics is mainly concerned with statistical consistency also known as calibration, which determines whether convergence to the minimizer of a surrogate loss implies convergence to the minimizer of the task performance measure as the sample size goes to infinity (Dembczyński et al. 2010; Nan et al. 2012; Gao and Zhou 2013; Zhao et al. 2013; Narasimhan et al. 2014; Koyejo et al. 2014, 2015). Here we give a stronger result which bounds the regret with respect to the performance metric by the regret with respect to the surrogate loss. Our result is valid for all finite sample sizes and informs about the rates of convergence.
We also note that two distinct frameworks are used to study the statistical consistency of classifiers with respect to performance metrics: Decision Theoretic Analysis (DTA), which assumes a test set of a fixed size, and Empirical Utility Maximization (EUM), in which the metric is defined by means of population quantities (Nan et al. 2012). In this context, our work falls into the EUM framework.
Parambath et al. (2014) presented an alternative approach to maximizing linear-fractional metrics by learning a sequence of binary classification problems with varying misclassification costs. While we were inspired by their theoretical analysis, their approach is, however, more complicated than the two-step approach analyzed here, which requires solving an ordinary binary classification problem only once. Moreover, as part of our proof, we show that by minimizing a strongly proper composite loss, we are implicitly minimizing cost-sensitive classification error for any misclassification costs without any overhead. Hence, the costs need not be known during learning, and can only be determined later on a separate validation sample by optimizing the threshold. Narasimhan et al. (2015) developed a general framework for designing provably consistent algorithms for complex multiclass performance measures. They relate the regret with respect to the target metric to the conditional probability estimation error measured in terms of \(L_1\)-metric. Their algorithms rely on using accurate class conditional probability estimates and multiple solving cost-sensitive multiclass classification problems.
The generalized performance metrics for binary classification are employed in the multilabel setting by means of one of the three averaging schemes (Waegeman et al. 2013; Parambath et al. 2014; Koyejo et al. 2015): instance-averaging (averaging errors over the labels, averaging metric over the examples), macro-averaging (averaging errors over the examples, averaging metric over the labels), and micro-averaging (averaging errors over the examples and the labels). Koyejo et al. (2015) characterize the optimal classifiers for multilabel metrics and prove the consistency of the plug-in method. Our regret bounds for multilabel classification can be seen as a follow up on their work.
1.2 Outline
The paper is organized as follows. In Sect. 2 we introduce basic concepts, definitions and notation. The main result is presented in Sect. 3 and proved in Sect. 4. Section 5 extends our results to the multilabel setting. The theoretical contribution of the paper is complemented by computational experiments in Sect. 6, prior to concluding with a summary in Sect. 7.
2 Problem setting
2.1 Binary classification
Some popular linear-fractional performance metrics expressed as functions of \({\mathrm {FN}}\) and \({\mathrm {FP}}\)
Metric | Expression |
---|---|
Accuracy | \(\mathrm {Acc} = 1 - {\mathrm {FN}}- {\mathrm {FP}}\) |
\(F_{\beta }\)-measure | \(F_{\beta } = \frac{(1+\beta ^2) (P - {\mathrm {FN}})}{(1+\beta ^2)P - {\mathrm {FN}}+ {\mathrm {FP}}}\) |
Jaccard similarity | \(J = \frac{P - {\mathrm {FN}}}{P + {\mathrm {FP}}}\) |
AM measure | \(\mathrm {AM} = \frac{2P(1-P) - P \,{\mathrm {FP}}- (1-P){\mathrm {FN}}}{2P(1-P)}\) |
Weighted accuracy | \(\mathrm {WA} = \frac{w_1(1-P) + w_2 P - w_1 {\mathrm {FP}}- w_2 {\mathrm {FN}}}{w_1(1-P) + w_2 P}\) |
2.2 Strongly proper composite losses
Here we briefly outline the theory of strongly proper composite loss functions. See Agarwal (2014) for a more detailed description.
Three popular strongly proper composite losses: squared-error, logistic and exponential losses
Loss function | Squared-error | Logistic | Exponential |
---|---|---|---|
\(\ell (y,f)\) | \((y-f)^2\) | \(\log \left( 1+e^{-fy}\right) \) | \(e^{-yf}\) |
\(c(1,\widehat{\eta })\) | \(4(1-\widehat{\eta })^2\) | \(-\log \widehat{\eta }\) | \(\sqrt{\frac{1-\widehat{\eta }}{\widehat{\eta }^{}}}\) |
\(c(-1,\widehat{\eta })\) | \(4\widehat{\eta }^2\) | \(-\log (1-\widehat{\eta })\) | \(\sqrt{\frac{\widehat{\eta }}{1-\widehat{\eta }^{}}}\) |
\(\psi (\widehat{\eta })\) | \(2 \widehat{\eta }-1\) | \(\log \frac{\widehat{\eta }}{1-\widehat{\eta }^{}}\) | \(\frac{1}{2} \log \frac{\widehat{\eta }}{1-\widehat{\eta }^{}}\) |
\(\lambda \) | 8 | 4 | 4 |
Table 2 presents some of the commonly used losses which are strongly proper composite. Note that the hinge loss\(\ell (y,f) = (1-yf)_+\), used, e.g., in support vector machines (Hastie et al. 2009), is not strongly proper composite (even not proper composite).
3 Main result
Lemma 1
The proof is quite long and hence is postponed to Sect. 4. Interestingly, the proof goes by an intermediate bound of the \(\varPsi \)-regret by a cost-sensitive classification regret. We note that the bound in Lemma 1 is in general unimprovable, in the sense that it is easy to find f, \(\varPsi \), \(\ell \), and distribution \(\Pr (x,y)\), for which the bound holds with equality (see proof for details). We split the constant in front of the bound into C and \(\lambda \), because C depends only on \(\varPsi \), while \(\lambda \) depends only on \(\ell \). Table 3 lists these constants for some popular metrics. We note that constant \(\gamma \) (lower bound on the denominator of \(\varPsi \)) will be distribution-dependent in general (as it can depend on \(P=\Pr (y=1)\)) and may not have a uniform lower bound which holds for all distributions.
Lemma 1 has the following interpretation. If we are able to find a function f with small \(\ell \)-regret, we are guaranteed that there exists a threshold \(\theta ^*\) such that \(h_{f,\theta ^*}\) has small \(\varPsi \)-regret. Note that the same threshold \(\theta ^*\) will work for any f, and the right hand side of the bound is independent of \(\theta ^*\). Hence, to minimize the right hand side we only need to minimize \(\ell \)-regret, and we can deal with the threshold afterwards.
Constants which appear in the bound of Lemma 1 for several performance metrics
Metric | \(\gamma \) | C |
---|---|---|
Accuracy | 1 | 2 |
\(F_{\beta }\)-measure | \(\beta ^2 P\) | \(\frac{1+\beta ^2}{\beta ^2 P}\) |
Jaccard similarity | P | \(\frac{J^*+1}{P}\) |
AM measure | \(2P(1-P)\) | \(\frac{1}{2P(1-P)}\) |
Weighted accuracy | \(w_1 P + w_2(1-P)\) | \(\frac{w_1+w_2}{w_1 P + w_2(1-P)}\) |
To make Lemma 1 easier to grasp, consider a special case when the performance metric \(\varPsi ({\mathrm {FP}},{\mathrm {FN}}) = 1 - {\mathrm {FP}}- {\mathrm {FN}}\) is the classification accuracy. In this case, (3) gives \(\varPsi ^{-1}(\theta ^*) = 1/2\). Hence, we obtained the well-known result that the classifier maximizing the accuracy is a threshold function on \(\eta \) at 1 / 2. Then, Lemma 1 states that given a real-valued f, we should take a classifier \(h_{f,\theta ^*}\) which thresholds f at \(\theta ^* = \psi (1/2)\). Using Table 2, one can easily verify that \(\theta ^* = 0\) for logistic, squared-error and exponential loss. This agrees with the common approach of thresholding the real-valued classifiers trained by minimizing these losses at 0 to obtain the label prediction. The bounds from the lemma are in this case identical (up to a multiplicative constant) to the bounds obtained by Bartlett et al. (2006).
Unfortunately, for more complicated performance metrics, the optimal threshold \(\theta ^*\) is unknown, as (3) contains unknown quantity \(\varPsi (h^*_{\varPsi })\), the value of the metric at optimum. The solution in this case is to, given f, directly search for a threshold which maximizes \(\varPsi (h_{f,\theta })\). This is the main result of the paper:
Theorem 1
Proof
The result follows immediately from Lemma 1: Solving \(\max _{\theta }\varPsi (h_{f,\theta })\) is equivalent to solving \(\min _\theta {\mathrm {Reg}}_{\varPsi }(h_{f,\theta })\), and \(\min _\theta {\mathrm {Reg}}_{\varPsi }(h_{f,\theta }) \le {\mathrm {Reg}}_{\varPsi }(h_{f,\theta ^*})\), where \(\theta ^*\) is the threshold given by Lemma 1. \(\square \)
- 1.
Find f with small \(\ell \)-regret, e.g., by using a learning algorithm minimizing \(\ell \)-risk on the training sample.
- 2.
Given f, solve \(\theta ^*_f = {\text {argmax}}_{\theta } \varPsi (h_{f,\theta })\).
In Theorem 2 below, we show that:Given f and validation sample \(\mathcal {T}\), solve \(\widehat{\theta }_f = {\text {argmax}}_{\theta } \widehat{\varPsi }(h_{f,\theta })\).
Theorem 2
Proof
We note that, contrary to a similar results by Koyejo et al. (2014), Theorem 2 does not require continuity of the cumulative distribution of \(\eta (x)\) around \(\theta ^*\).
4 Proof of Lemma 1
The proof can be skipped without affecting the flow of later sections. The proof consists of two steps. First, we bound the \(\varPsi \)-regret of any classifier h by its cost-sensitive classification regret (introduced below). Next, we show that there exists a threshold \(\theta ^*\), such that for any f, the cost-sensitive classification regret of \(h_{f,\theta ^*}\) is upperbounded by the \(\ell \)-regret of f. These two steps will be formalized as Proposition 1 and Proposition 2.
Proposition 1
Proof
Proposition 2
Proof
First, we will show that (7) holds conditionally for every x. To this end, we fix x and deal with \(h(x) \in \{-1,1\}\), \(f(x) \in {\mathbb {R}}\) and \(\eta (x) \in [0,1]\), using a shorthand notation \(h,f,\eta \).
We note that derivation of (9) follows the steps of the proof of Lemma 4 in Menon et al. (2013), while (10) and (11) were shown in the proof of Theorem 13 by Agarwal (2014). Hence, the proof is essentially a combination of existing results, which are rederived here for the sake of completeness. \(\square \)
Note that the proof actually specifies the exact value of the universal threshold, \(\theta ^* = \psi (\alpha )\), where \(\alpha \) is given by (5).
5 Multilabel classification
What follows next depends on the way in which the binary classification performance metric is applied in the multilabel setting. We consider two ways of turning binary classification metric into multilabel metric: the macro-averaging and the micro-averaging (Manning et al. 2008; Parambath et al. 2014; Koyejo et al. 2015).
5.1 Macro-averaging
Theorem 3
Proof
The theorem follows from applying Theorem 1 once for each label, and then averaging the bounds over the labels. \(\square \)
Theorem 3 suggests a straightforward decomposition into m independent binary classification problems, one for each label \(y_1,\ldots ,y_m\), and running (independently for each problem) the two-step procedure described in Sect. 3: For \(i=1,\ldots ,m\), we learn a function \(f_i\) with small \(\ell \)-regret with respect to label \(y_i\), and tune the threshold \(\theta ^*_{f_i}\) to optimize \(\varPsi (h_{f_i,\theta })\) (similarly as in the binary classification case, one can show that tuning the threshold on a separate validation sample is sufficient). Due to decomposition of \(\varPsi _{\mathrm {macro}}\) into the sum over the labels, this simple procedure turns out to be sufficient. As we shall see, the case of micro-averaging becomes more interesting.
5.2 Micro-averaging
Theorem 4
Proof
Theorem 4 suggests a decomposition into m independent binary classification problems, one for each label \(y_1,\ldots ,y_m\), and training m real-valued classifiers \(f_1,\ldots ,f_m\) with small \(\ell \)-regret on the corresponding label. Then, however, contrary to macro-averaging, a single threshold, shared among all labels, is tuned by optimizing \(\varPsi _{\mathrm {micro}}\) on a separate validation sample.
6 Empirical results
We perform experiments on synthetic and benchmark data to empirically study the two-step procedure analyzed in the previous sections. To this end, we minimize a surrogate loss in the first step to obtain a real-valued function f, and in the second step, we tune a threshold \(\hat{\theta }\) on a separate validation set to optimize a given performance metric. We use logistic loss in this procedure as a surrogate loss. Recall that logistic loss is 4-strongly proper composite (see Table 2). We compare its performance with hinge loss, which is even not a proper composite function.
As our task performance metrics, we take the F-measure (\(F_{\beta }\)-measure with \(\beta =1\)) and the AM measure (which is a special case of Weighted Accuracy with weights \(w_1 = P\) and \(w_2 = 1-P\)). We could also use the Jaccard similarity coefficient; it turns out, however, that the threshold optimized for the F-measure coincides with the optimal threshold for the Jaccard similarity coefficient (this is because the Jaccard similarity coefficient is strictly monotonic in the F-measure and vice versa), so the latter measure does not give anything substantially different than the F-measure.
The experiments on benchmark data are split into two parts. The first part concerns binary classification problems, while the second part multi-label classification.
The purpose of this study is not about comparing the two-step approach with alternative methods; this has already been done in the previous work on the subject, see, e.g., Nan et al. (2012) and Parambath et al. (2014). We also note that similar experiments have been performed in the cited papers on the statistical consistency of generalized performance metrics (Koyejo et al. 2014; Narasimhan et al. 2014; Parambath et al. 2014; Koyejo et al. 2015). Therefore, we unavoidably repeat some of the results obtained therein, but the main novelty of the experiments reported here is that we emphasize the difference between strongly proper composite losses and non-proper losses.
6.1 Synthetic data
We performed two experiments on synthetic data. The first experiment deals with a discrete domain in which we learn within a class of all possible classifiers. The second experiment concerns continuous domain in which we learn within a restricted class of linear functions.
First experiment We let the input domain X to be a finite set, consisting of 25 elements, \(X=\{1,2,\ldots ,25\}\), and take \(\Pr (x)\) to be uniform over X, i.e., \(\Pr (x=i)=1/25\). For each \(x \in X\), we randomly draw a value of \(\eta (x)\) from the uniform distribution on the interval [0, 1]. In the first step, we take an algorithm which minimizes a given surrogate loss \(\ell \) within the class of all function \(f :X \rightarrow {\mathbb {R}}\). Hence, given the training data of size n, the algorithm computes the empirical minimizer of surrogate loss \(\ell \) independently for each x. As surrogate losses, we use logistic and hinge loss. In the second step, we tune the threshold \(\hat{\theta }\) on a separate validation set, also of size n. For each n, we repeat the procedure 100,000 times, averaging over samples and over models (different random choices of \(\eta (x)\)). We start with \(n=100\) and increase the number of training examples up to \(n=10{,}000\). The \(\ell \)-regret and \(\varPsi \)-regret can be easily computed, as the distribution is known and X is discrete.
Second experiment We take \(X = {\mathbb {R}}^2\) and generate \(x \in X\) from a standard Gaussian distribution. We use a logistic model of the form \(\eta (x) = \frac{1}{1 + \exp (-a_0 - a^\top x)}\). The weights \(a = (a_1, a_2)\) and \(a_0\) are also drawn from a standard Gaussian. For a given model (set of weights), we take training sets of increasing size from \(n=100\) up to \(n=3000\), using 20 different sets for each n. We also generate one test set of size 100,000. For each n, we use 2/3 of the training data to learn a linear model \(f(x) = w_0 + w^\top x\), using either support vector machines (SVM, with linear kernel) or logistic regression (LR). We use implementation of these algorithms from the LibLinear package (Fan et al. 2008).^{4} The remaining 1/3 of the training data is used for tuning the threshold. We average the results over 20 different models.
Basic statistics for binary classification benchmark datasets
Dataset | No. of examples | No. of features |
---|---|---|
covtype | 581,012 | 54 |
gisette | 7000 | 5000 |
6.2 Benchmark data for binary classification
Basic statistics for multi-label benchmark data sets
Data set | No. of labels | No. of training examples | No. of test examples | No. of features |
---|---|---|---|---|
scene | 6 | 1211 | 1169 | 294 |
yeast | 14 | 1500 | 917 | 103 |
mediamill | 101 | 30,993 | 12,914 | 120 |
6.3 Benchmark data for multi-label classification
In the last experiment we use three multi-label benchmark data sets.^{6} Table 5 provides a summary of basic statistics of these datasets. The aim of the experiment is to verify the theoretical results in Sect. 5 on learning the micro- and macro-averaged performance metrics. We use the F-measure and the AM-measure as in previous experiments.
The data sets are already split into the training and testing parts. As before we train a linear model using either SVM or LR on 2/3 of training examples. The rest of training data is used for tuning the threshold. For optimizing macro-averaged measures, we tune the threshold separately for each label. This approach agrees with our analysis given in Sect. 5.1. For micro-averaging, we tune a common threshold for all labels: we simply collect predictions for all labels and find the best threshold using these values. This approach is justified by the theoretical analysis in Sect. 5.2. Hence, the only difference between micro- and macro-versions of the algorithms is whether a single or multiple thresholds are tuned. In total we use 8 algorithms: two learning algorithms (LR/SVM), two performance measures (F/AM), and two types of averaging (Macro/Micro). Note that our experiments include evaluating algorithms tuned for macro-averaging in terms of micro-averaged metrics, and vice versa. The goal of such cross-analysis is to determine the impact of threshold sharing for both averaging schemes. As before, we incrementally increase the size of the training set and repeat training and threshold tuning 20 times (we use random draws of training instances into the proper training and the validation parts; the test set is always the same, as originally specified for each data set). The results are given in Fig. 4.
The plots generally agree with the conclusions coming from the theoretical analysis, with some intriguing exceptions, however. As expected, LR tuned for a given performance metric gets the best result with respect to that metric in most of the cases. For the scene data set, however, the methods tuned for the micro-averaged metrics (single threshold shared among labels) outperform the ones tuned for macro-averaged metrics (separate thresholds tuned for each label), even when evaluated in terms of macro-averaged metrics. A similar result has been obtained by Koyejo et al. (2015). It seems that tuning a single threshold shared among all labels can lead to a more stable solution that is less prone to overfitting, even though it is not the optimal thing to do for macro-averaged measures. We further report that, interestingly, SVM outperform LR in terms of Macro-F on mediamill and this is the only case in which SVM get a better result than LR.
7 Summary
We present a theoretical analysis of a two-step approach to optimize classification performance metrics, which first learns a real-valued function f on a training sample by minimizing a surrogate loss, and then tunes the threshold on f by optimizing the target performance metric on a separate validation sample. We show that if the metric is a linear-fractional function, and the surrogate loss is strongly proper composite, then the regret of the resulting classifier (obtained from thresholding real-valued f) measured with respect to the target metric is upperbounded by the regret of f measured with respect to the surrogate loss. The proof of our result goes by an intermediate bound of the regret with respect to the target measure by a cost-sensitive classification regret. As a byproduct, we get a bound on the cost-sensitive classification regret by a surrogate regret of a real-valued function which holds simultaneously for all misclassification costs. We also extend our results to cover multilabel classification and provide regret bounds for micro- and macro-averaging measures. Our findings are backed up in a computational study on both synthetic and real data sets.
Footnotes
- 1.
Throughout the paper, we follow the convention that all conditional quantities are lowercase (regret, risk), while all unconditional quantities are uppercase (Regret, Risk).
- 2.
- 3.
The fact that a single threshold is sufficient for consistency of micro-averaged performance measures was already noticed by Koyejo et al. (2015).
- 4.
Software available at http://www.csie.ntu.edu.tw/~cjlin/liblinear.
- 5.
Datasets are taken from LibSVM repository: http://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets.
- 6.
Datasets are taken from LibSVM repository: http://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets.
Notes
Acknowledgments
W. Kotłowski has been supported by the Polish National Science Centre under Grant No. 2013/11/D/ST6/03050. K. Dembczyński has been supported by the Polish National Science Centre under Grant No. 2013/09/D/ST6/03917.
References
- Agarwal, S. (2014). Surrogate regret bounds for bipartite ranking via strongly proper losses. Journal of Machine Learning Research, 15, 1653–1674.MathSciNetMATHGoogle Scholar
- Bartlett, P. L., Jordan, M. I., & McAuliffe, J. D. (2006). Convexity, classification, and risk bounds. Journal of the American Statistical Association, 101(473), 138–156.MathSciNetCrossRefMATHGoogle Scholar
- Dembczyński, K., Cheng, W., & Hüllermeier, E. (2010). Bayes optimal multilabel classification via probabilistic classifier chains. In ICML 2010 (pp. 279–286). Omnipress.Google Scholar
- Dembczyński, K., Waegeman, W., Cheng, W., & Hüllermeier, E. (2012). On loss minimization and label dependence in multi-label classification. Machine Learning, 88, 5–45.MathSciNetCrossRefMATHGoogle Scholar
- Dembczyński, K., Jachnik, A., Kotłowski, W., Waegeman, W., & Hüllermeier, E. (2013). Optimizing the f-measure in multi-label classification: Plug-in rule approach versus structured loss minimization. In ICML.Google Scholar
- Devroye, L., Györfi, L., & Lugosi, G. (1996). A probabilistic theory of pattern recognition (1st ed.). Berlin: Springer.CrossRefMATHGoogle Scholar
- Fan, R. E., Chang, K. W., Hsieh, C. J., Wang, X. R., & Lin, C. J. (2008). LIBLINEAR: A library for large linear classification. Journal of Machine Learning Research, 9, 1871–1874.MATHGoogle Scholar
- Gao, W., & Zhou, Z. H. (2013). On the consistency of multi-label learning. Artificial Intelligence, 199–200, 22–44.MathSciNetCrossRefMATHGoogle Scholar
- Hastie, T., Tibshirani, R., & Friedman, J. H. (2009). Elements of statistical learning: Data mining, inference, and prediction. Berlin: Springer.CrossRefMATHGoogle Scholar
- Jansche, M. (2005). Maximum expected F-measure training of logistic regression models. In HLT/EMNLP 2005 (pp. 736–743).Google Scholar
- Jansche, M. (2007). A maximum expected utility framework for binary sequence labeling. In ACL 2007 (pp. 736–743).Google Scholar
- Koyejo, O., Natarajan, N., Ravikumar, PK., & Dhillon, IS. (2014). Consistent binary classification with generalized performance metrics. In Neural information processing systems (NIPS).Google Scholar
- Koyejo, O., Natarajan, N., Ravikumar, P., & Dhillon, IS. (2015). Consistent multilabel classification. In Neural information processing systems (NIPS).Google Scholar
- Lewis, D. (1995). Evaluating and optimizing autonomous text classification systems. In SIGIR 1995 (pp. 246–254).Google Scholar
- Manning, C. D., Raghavan, P., & Schütze, H. (2008). Introduction to information retrieval. Cambridge: Cambridge University Press.CrossRefMATHGoogle Scholar
- Menon, A. K., Narasimhan, H., Agarwal, S., & Chawla, S. (2013). On the statistical consistency of algorithms for binary classification under class imbalance. In International conference on machine learning (ICML).Google Scholar
- Musicant, D. R., Kumar, V., & Ozgur, A. (2003). Optimizing f-measure with support vector machines. In FLAIRS conference (pp. 356–360)Google Scholar
- Nan, Y., Chai, K. M. A., Lee, WS., & Chieu, H.L. (2012). Optimizing F-measure: A tale of two approaches. In International conference on machine learning (ICML).Google Scholar
- Narasimhan, H., Vaish, R., & Agarwal, S. (2014). On the statistical consistency of plug-in classifiers for non-decomposable performance measures. In Neural information processing systems (NIPS).Google Scholar
- Narasimhan, H., Ramaswamym, H. G., Saha, A., & Agarwal, S. (2015). Consistent multiclass algorithms for complex performance measures. In International conference on machine learning (ICML).Google Scholar
- Parambath, S. P., Usunier, N., & Grandvalet, Y. (2014). Optimizing F-measures by cost-sensitive classification. In Neural information processing systems (NIPS).Google Scholar
- Petterson, J., & Caetano, T. S. (2010). Reverse multi-label learning. Advances in Neural Information Processing Systems, 24, 1912–1920.Google Scholar
- Petterson, J., & Caetano, T. S. (2011). Submodular multi-label learning. Advances in Neural Information Processing Systems, 24, 1512–1520.Google Scholar
- Reid, M. D., & Williamson, R. C. (2010). Composite binary losses. Journal of Machine Learning Research, 11, 2387–2422.MathSciNetMATHGoogle Scholar
- Reid, M. D., & Williamson, R. C. (2011). Information, divergence and risk for binary experiments. Journal of Machine Learning Research, 12, 731–817.MathSciNetMATHGoogle Scholar
- Tsochantaridis, I., Joachims, T., Hofmann, T., & Altun, Y. (2005). Large margin methods for structured and interdependent output variables. Journal of Machine Learning Research, 6, 1453–1484.MathSciNetMATHGoogle Scholar
- Waegeman, W., Dembczyński, K., Jachnik, A., Cheng, W., & Hüllermeier, E. (2013). On the Bayes-optimality of F-measure maximizers. Journal of Machine Learning Research, 15, 3513–3568.MATHGoogle Scholar
- Zhao, M. J., Edakunni, N., Pocock, A., & Brown, G. (2013). Beyond Fano’s inequality: Bounds on the optimal F-score, BER, and cost-sensitive risk and their implications. Journal of Machine Learning Research, 14, 1033–1090.MathSciNetMATHGoogle Scholar
Copyright information
Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.