Primers • Evaluation Metrics, ROCCurves and Imbalanced Datasets
 Introduction
 Evaluation Metrics for the Classification problem
 Types of prediction errors
 Accuracy
 Confusion Matrix
 Precision and Recall
 \(F_1\) score
 Sensitivity and Specificity
 Putting it together: Precision, Recall/Sensitivity, Specificity, and NPV
 Calculating precision, sensitivity and specificity
 Receiver Operating Characteristic (ROC) Curve
 Detection error tradeoff (DET) curve
 Example Walkthrough with Code
 Evaluation Metrics for the Regression Problem
 Evaluation Metrics for NLP/Speech Models
 Evaluation Metrics for Compression Models
 Evaluation Metrics for GANbased Models
 Evaluation Metrics for recommender systems
 Further reading
 References
 Citation
Introduction
 Deep learning tasks can be complex and hard to measure: how do we know whether one network is better than another? In some simpler cases such as regression, the loss function used to train a network can be a good measurement of the network’s performance.
 However, for many realworld tasks, there are evaluation metrics that encapsulate, in a single number, how well a network is doing in terms of real world performance. These evaluation metrics allow us to quickly see the quality of a model, and easily compare different models on the same tasks.
 In this section, we review and compare some of the popular evaluation metrics typically used for classification tasks, and how they should be used depending on the the dataset. Next, we also go over how one can tune the probability thresholds for the particularly metrics. Finally, we’ll go through some case studies of different tasks and their metrics.
Evaluation Metrics for the Classification problem
 Let’s consider a simple binary classification problem, where we are trying to predict if a patient is healthy or has pneumonia. We have a test set with 10 patients, where 9 patients are healthy (shown as green squares) and 1 patient has pneumonia (shown as a red square). The ground truth for your test set is shown below:
 We’ve trained three models for this task (Model1, Model2, Model3), and we’d like to compare the performance of these models. The predictions from each model on the test set are shown below:
Types of prediction errors
 When making a prediction for a twoclass classification problem, the following types of errors can be made by a classifier:
 False Positive (FP): predict an event when there was no event. This is also referred to as a Type I error in statistical hypothesis testing (defined as the mistaken rejection of an actually true null hypothesis).
 False Negative (FN): predict no event when in fact there was an event. This is also referred to as a Type II error in statistical hypothesis testing (defined as the failure to reject a null hypothesis that is actually false).
 True Positive (TP): predict an event when there was an event.
 True Negative (TN): predict no event when in fact there was no event.
 In general, the error type can be interpreted as follows:
 The first word indicates the prediction outcome. If our prediction was correct, it’s true, else it’s false.
 The second word indicates the actual prediction. If our prediction was that an event occurred, it’s positive, else it’s negative.
Accuracy

To compare models, we could first use accuracy, which is the fraction of the number of correctly classified examples relative to the total number of examples:
\[\text{Accuracy} = \frac{\sum_{x_i \in X_{test}} \mathbb{1}\{f(x_i) = y_i\}}{\mid X_{test} \mid}\] For instance, if the classifier is 90% correct, it means that out of 100 instances, it correctly predicts the class for 90 of them.

If we use accuracy as your evaluation metric, it seems that the best model is Model1.
 However, accuracy can be misleading if the number of samples per class in the task at hand is unbalanced. Having a dataset with two classes only, where the first class is 90% of the data, and the second completes the remaining 10%. If the classifier predicts every sample as belonging to the first class, the accuracy reported will be of 90% but this classifier is in practice useless. In general, when you have class imbalance (which is most of the time!), accuracy is not a good metric to use. With imbalanced classes, it’s easy to get a high accuracy without actually making useful predictions. So, accuracy as an evaluation metrics makes sense only if the class labels are uniformly distributed.
Confusion Matrix
 Accuracy doesn’t discriminate between errors (i.e., it treats misclassifying a patient with pneumonia as healthy the same as misclassifying a healthy patient as having pneumonia). A confusion matrix is a tabular format for showing a more detailed breakdown of a model’s correct and incorrect classifications.
 A confusion matrix for binary classification is shown below:
Precision and Recall
 In case of an imbalanced dataset scenario (where you have an abundance of negatives and a dearth of positives), precision and recall are appropriate performance metrics. Both precision and recall are focused on the positive class (the minority class) and are unconcerned with the true negatives (majority class). Put simply, precision and recall are the preferred metrics in case of a class imbalance scenario when you have a lot of negatives and a few positives, for e.g., detecting the wakeword in a typical voice assistant pipeline. In other words, precision and recall make it possible to assess the performance of a classifier on the minority class.
 Precision is defined as the fraction of relevant instances among all retrieved instances.
 Recall, sometimes referred to as sensitivity, is the fraction of retrieved instances among all relevant instances.
 Note that precision and recall are computed for each class. They are commonly used to evaluate the performance of classification or information retrieval systems.
A perfect classifier has precision and recall both equal to 1.
 It is often possible to calibrate the number of results returned by a model and improve precision at the expense of recall, or vice versa.
 Precision and recall should always be reported together and are not quoted individually. This is because it is easy to vary the sensitivity of a model to improve precision at the expense of recall, or vice versa.
 As an example, imagine that the manufacturer of a pregnancy test needed to reach a certain level of precision, or of specificity, for FDA approval. The pregnancy test shows one line if it is moderately confident of the pregnancy, and a double line if it is very sure. If the manufacturer decides to only count the double lines as positives, the test will return far fewer positives overall, but the precision will improve, while the recall will go down. This shows why precision and recall should always be reported together.
 The figure below (taken from the Wikipedia article on precision and recall) shows a graphical representation of precision and recall:
 Formally, precision and recall can be defined as:
 Precision: Out of all the samples marked positive, how many were actually positive (i.e., the true positives)?
 Recall: Out of all the samples that are actually positive, how many were marked positive (i.e., the true positives)?
 From the above definitions it is clear that with PR metrics, the focus is on the positive class (also called the relevant class).
 Precision and recall are typically juxtaposed together when reported.

In the section on PrecisionRecall (PR) Curves, we explore how to get the best out of these two metrics using PR curves.
 Key takeaways
 Precision: how many selected items are relevant?
 Recall: how many relevant items are selected?
Memory map for precision/recall formulae
 The formula for precision follows the tripleP rule: calculation of Precision related to terms of True Positive and False Positive, which is one half of the confusion matrix.
 The formula for recall is related to the first row of terms in the confusion matrix.
Historical Background
 This section is optional and offers a historical walkthrough of how precision, recall and F1score came about, so you may skip to the next section if so desired.
 Precision and recall were first defined by the American scientist Allen Kent and his colleagues in their 1955 paper Machine literature searching VIII. Operational criteria for designing information retrieval systems.
 Kent served in the US Army Air Corps in World War II, and was assigned after the war by the US military to a classified project at MIT in mechanized document encoding and search.
 In 1955, Kent and his colleagues Madeline Berry, Fred Luehrs, and J.W. Perry were working on a project in information retrieval using punch cards and reeltoreel tapes. The team found a need to be able to quantify the performance of an information retrieval system objectively, allowing improvements in a system to be measured consistently, and so they published their definition of precision and recall.
 They described their ideas as a theory underlying the field of information retrieval, just as the second law of thermodynamics “underlies the design of a steam engine, regardless of its type or power rating”.
 Since then, the definitions of precision and recall have remained fundamentally the same, although for search engines the definitions have been modified to take into account certain nuances of human behavior, giving rise to the modified metrics precision @ \(k\) and mean average precision (mAP), which are the values normally quoted in information retrieval contexts today.
 In 1979, the Dutch computer science professor Cornelis Joost van Rijsbergen recognized the problems of defining search engine performance in terms of two numbers and decided on a convenient scalar function that combines the two. He called this metric the Effectiveness function and assigned it the letter E. This was later modified to the \(F_1\) score, or \(F_{\beta}\) score, which is still used today to summarize precision and recall.
Examples
 Precision and recall can be best explained using examples. Consider the case of evaluating how well does a robot sifts good apples from rotten apples. A robot looks into the basket and picks out all the good apples, leaving the rotten apples behind, but is not perfect and could sometimes mistake a rotten apple for a good apple orange.
 After the robot finishes picking the good apples, precision and recall can be calculated as:
 Precision: number of good apples picked out of all the picked apples.
 Recall: number of good apples picked out of all possible good apples.
 Precision is about exactness, classifying only one instance correctly yields 100% precision, but a very low recall, it tells us how well the system identifies samples from a given class.
 Recall is about completeness, classifying all instances as positive yields 100% recall, but a very low precision, it tells how well the system does and identify all the samples from a given class.
 As another example, consider the task of information retrieval. As such, precision and recall can be calculated as:
 Precision: number of relevant documents retrieved out of all retrieved documents.
 Recall: number of relevant documents retrieved out of all relevant documents.
Applications
 Precision and recall are measured for every possible class in your dataset. So, precision and recall metrics are relatively much more appropriate (especially compared to accuracy) when dealing with imbalanced classes.
An important point to note is that PR are able to handle class imbalance in scenarios where the positive class (also called the minority class) is rare. If, however, the dataset is imbalanced in such a way that the negative class is the one that’s rare, PR curves are suboptimal and can be misleading. In these cases, ROC curves might be a better fit.
 So when do we use PR metrics? Here’s the typical usecases:
 When two classes are equally important: PR would be the metrics to use if the goal of the model is to perform equally well on both classes. Image classification between cats and dogs is a good example because the performance on cats is equally important on dogs.
 When minority class is more important: PR would be the metrics to use if the focus of the model is to identify correctly as many positive samples as possible. Take spam detectors for example, the goal is to find all the possible spam emails. Regular emails are not of interest at all — they overshadow the number of positives.
Formulae

Mathematically, precision and recall are defined as,
\[\operatorname {Precision}=\frac{TP}{TP + FP} \\\] \[\operatorname{Recall}=\frac{TP}{TP + FN} \\\] where,
 \(TP\) is the True Positive Rate, i.e., the number of instances which are relevant and which the model correctly identified as relevant.
 \(FP\) is the False Positive Rate, i.e., the number of instances which are not relevant but which the model incorrectly identified as relevant.
 \(FN\) is the false negative rate, i.e., the number of instances which are relevant and which the model incorrectly identified as not relevant.
 where,

In the content of the robot sifting good apples from the rotten ones,
\[\operatorname {Precision}=\frac{\text { # of picked good apples }}{\text { # of picked apples }}\] \[\operatorname{Recall}=\frac{\text { # of picked good apples }}{\text { # of good apples }}\] 
In the context of information retrieval,
\[\operatorname {Precision}=\frac{\text { retrieved relevant documents }}{\text { all retrieved documents }}\] \[\operatorname{Recall}=\frac{\text { retrieved relevant documents }}{\text { all relevant documents }}\]
Precision/Recall Tradeoff
 Because high precision and high recall are what every model optimizes for. However, depending on the problem at hand, you either care about high precision or high recall.
 Examples of high precision:
 For a model that detects shop lifters, the focus should be on developing a high precision model by reducing false positives (note that precision is given by \(\frac{TP}{TP+FP}\) and since \(FP\) features in the denominator, reducing \(FP\) leads to high precision). This implies that if we tag someone as a shop lifter, we’d like to make sure we do so with high confidence.
 Examples of high recall:
 In an adult content detection problem, the focus should be on developing a high recall model by reducing false negatives (note that recall is given by \(\frac{TP}{TP+FN}\) and since \(FN\) features in the denominator, reducing \(FN\) leads to high recall). This implies that if the model classified a video as good for kids (i.e., not having adult content), it should be marked so with high confidence.
 In a disease detection scenario, the focus should be on developing a high recall model by reducing false negatives. This implies that if the model classified a patient as not having the disease, it should be done do with high confidence else it can prove fatal.
 In an autonomous car driving scenario, the focus should be on developing a high recall model by reducing false negatives. This implies that if the model determined that there was no obstacle in the car’s surrounding radius, it should be done do with high confidence else fatalities can occur.
 Often, there is an inverse relationship between precision and recall, where it is possible to increase one at the cost of reducing the other. This is called the precision/recall tradeoff. However, in some scenarios, it is important to strike the right balance between both:
 As an example (from the Wikipedia article on Precision and Recall), brain surgery provides an illustrative example of the tradeoff. Consider a brain surgeon removing a cancerous tumor from a patient’s brain. The surgeon needs to remove all of the tumour cells since any remaining cancer cells will regenerate the tumor. Conversely, the surgeon must not remove healthy brain cells since that would leave the patient with impaired brain function. The surgeon may be more liberal in the area of the brain he removes to ensure he has extracted all the cancer cells. This decision increases recall but reduces precision. On the other hand, the surgeon may be more conservative in the brain he removes to ensure he extracts only cancer cells. This decision increases precision but reduces recall. That is to say, greater recall increases the chances of removing healthy cells (negative outcome) and increases the chances of removing all cancer cells (positive outcome). Greater precision decreases the chances of removing healthy cells (positive outcome) but also decreases the chances of removing all cancer cells (negative outcome).
 The following diagram shows the PrecisionRecall tradeoff. As we increase the recall rate by adjusting the classification threshold of a model, the precision rate is decreased and vice versa.
Case studies
Disease diagnosis
 Consider our classification problem of pneumonia detection. It is crucial that we find all the patients that are suffering from pneumonia. Predicting patients with pneumonia as healthy is not acceptable (since the patients will be left untreated).
 Thus, a natural question to ask when evaluating our models is: Out of all the patients with pneumonia, how many did the model predict as having pneumonia? The answer to this question is given by the recall.
 The recall for each model is given by:
 Imagine that the treatment for pneumonia is very costly and therefore you would also like to make sure only patients with pneumonia receive treatment.
 A natural question to ask would be: Out of all the patients that are predicted to have pneumonia, how many actually have pneumonia? This metric is the precision.
 The precision for each model is given by:
Search engine
 Imagine that you are searching for information about cats on your favorite search engine. You type ‘cat’ into the search bar.
 The search engine finds four web pages for you. Three pages are about cats, the topic of interest, and one page is about something entirely different, and the search engine gave it to you by mistake. In addition, there are four relevant documents on the internet, which the search engine missed.

In this case we have three true positives, so \(TP=3\). There is one false positive, \(FP=1\). And there are four false negatives, so \(FN=4\). Note that to calculate precision and recall, we do not need to know the total number of true negatives (the irrelevant documents which were not retrieved).

The precision is given by,
 While the recall is given by,
PrecisionRecall Curve
 In the section on Precision and Recall, we see that when a dataset has imbalanced classes, precision and recall are better metrics than accuracy. Similarly, for imbalanced classes, a PrecisionRecall curve is more suitable than a ROC curve.
 A PrecisionRecall curve is a plot of the Precision (yaxis) and the Recall (xaxis) for different thresholds, much like the ROC curve. Note that in computing precision and recall there is never a use of the true negatives, these measures only consider correct predictions.
Area Under the PR Curve (AUPRC)
 Similar to the AUROC, the AUPRC summarizes the curve with a range of threshold values as a single score.
 The score can then be used as a point of comparison between different models on a binary classification problem where a score of 1.0 represents a model with perfect skill.
Key takeaways: Precision, Recall and ROC/PR Curves
 ROC Curve: summaries the tradeoff between the True Positive Rate and False Positive Rate for a predictive model using different probability thresholds.
 PrecisionRecall Curve: summaries the tradeoff between the True Positive Rate and the positive predictive value for a predictive model using different probability thresholds.
 In the same way it is better to rely on precision and recall rather than accuracy in an imbalanced dataset scenario (since it can offer you an incorrect picture of the classifier’s performance), a PrecisionRecall curve is better to calibrate the probability threshold compared to the ROC curve. In other words, ROC curves are appropriate when the observations are balanced between each class, whereas precisionrecall curves are appropriate for imbalanced datasets. In both cases, the area under the curve (AUC) can be used as a summary of the model performance.
Metric  Formula  Description 

Accuracy  $$\frac{TP+TN}{TP+TN + FP+FN}$$  Overall performance of model 
Precision  $$\frac{TP}{TP + FP}$$  How accurate the positive predictions are 
Recall/Sensitivity  $$\frac{TP}{TP + FN}$$  Coverage of actual positive sample 
Specificity  $$\frac{TN}{TN + FP}$$  Coverage of actual negative sample 
F1score  $$2 \times\frac{\textrm{Precision} \times \textrm{Recall}}{\textrm{Precision} + \textrm{Recall}}$$  Harmonic mean of Precision and Recall 
\(F_1\) score
 Precision and recall are both useful, but having multiple evaluation metrics makes it difficult to directly compare models. From Andrew Ng’s machine learning book:
“Having multiplenumber evaluation metrics makes it harder to compare algorithms. Better to combine them to a single evaluation metric. Having a singlenumber evaluation metric speeds up your ability to make a decision when you are selecting among a large number of classifiers. It gives a clear preference ranking among all of them, and therefore a clear direction for progress.”  Machine Learning Yearning
 Furthermore, considering the Precision Recall tradeoff, a balancing metric that combines the two terms is helpful. This metric is called the F1 score. Mathematically,
 Typically precision and recall are combined together into a single number metric called the \(F_1\) score, which is the harmonic mean of the precision and recall. The \(F_1\) score eases comparison of different systems, and problems with many classes. It is mathematically defined as:
 If we consider either precision or recall to be more important than the other, then we can use the \(F_{\beta}\) score, which is a weighted harmonic mean of precision and recall. This is useful, for example, in the case of a medical test, where a false negative may be extremely costly compared to a false positive. The \(F_{\beta}\) score formula is more complex:
 More on the FBeta score here.
Calculating \(F_1\) score
 Imagine that we consider precision and recall to be of equal importance for our purposes. In this case, we will use the \(F_1\)score to summarize precision and recall together.
 For the above example of disease diagnosis, let’s calculate the \(F_1\) score for each model based on the numbers for precision and recall,
 For the above example of a search engine, let’s plug in the numbers for precision and recall into the formula for \(F_1\)score,
 Note that the \(F_1\) score of 0.55 lies between the precision and recall values of 0.75 and 0.43 respectively. This illustrates how the \(F_1\) score can be a convenient way of averaging the precision and recall in order to condense them into a single number.
Sensitivity and Specificity
 When we need to express model performance in two numbers, an alternative twonumber metric to precision and recall is sensitivity and specificity. This is commonly used for medical stated sensitivity and specificity for a device or testing kit printed on the side of the box, or in the instruction leaflet.
 Sensitivity and specificity can be defined as follows:
 Sensitivity: can be thought of as the extent to which actual positives are not overlooked, so false negatives are few. Note that sensitivity is the same as recall.
 Specificity: also called the true negative rate, measures the proportion of actual negatives that are correctly identified as such, i.e., is the extent to which actual negatives are classified as such (so false positives are few).
 Mathematically, sensitivity and specificity can be defined as:
 In the context of identifying the number of people with a disease,
 Sensitivity therefore quantifies the avoiding of false negatives, and specificity does the same for false positives.
 Specificity also uses \(TN\), the number of true negatives. This means that sensitivity and specificity use all four numbers in the confusion matrix, as opposed to precision and recall which only use three.
 The number of true negatives corresponds to the number of patients identified by the test as having the disease when they did not have the disease, or alternatively the number of irrelevant documents which the search engine did not retrieve.
 Taking a probabilistic interpretation, we can view specificity as the probability of a negative test given that the patient is well, while the sensitivity is the probability of a positive test given that the patient has the disease.
 The figure below (taken from the Wikipedia article on sensitivity and specificity) shows a graphical representation of sensitivity and specificity:
 Key takeaways
 Sensitivity: how many relevant items are selected?
 Specificity: how many negative selected elements items are truly negative?
Precision and Recall vs. Sensitivity and Specificity
 Sensitivity and specificity are preferred to precision and recall in the medical domain, while precision and recall are the most commonly used metrics for information retrieval. This initially seems strange, since both pairs of metrics are measuring the same thing: the performance of a binary classifier.
 The reason for this discrepancy is that when we are measuring the performance of a search engine, we only care about the returned results, so both precision and recall are measured in terms of the true and false positives. However, if we are testing a medical device, it is important to take into account the number of true negatives, since these represent the large number of patients who do not have the disease and were correctly categorized by the device.
 In medical context, here’s scenarios where focusing on one of these two might be important:
 Sensitivity: the percentage of sick people who are correctly identified as having the condition.
 Specificity: the percentage of healthy people who are correctly identified as not having the condition.
Putting it together: Precision, Recall/Sensitivity, Specificity, and NPV
 The following table packs all of the metrics that we discussed above:
Calculating precision, sensitivity and specificity
 Let’s calculate the precision, sensitivity and specificity for the below example of disease diagnosis.
 Suppose we have a medical test which is able to identify patients with a certain disease.
 We test 20 patients and the test identifies 8 of them as having the disease.
 Of the 8 identified by the test, 5 actually had the disease (true positives), while the other 3 did not (false positives).
 We later find out that the test missed 4 additional patients who turned out to really have the disease (false negatives).
 We can represent the 20 patients using the following confusion matrix:
True state of patient's health  

Disease  No disease  
Test result  Alert  5  3 
No alert  4  8 

The relevant values for calculating precision and recall are \(TP=5\), \(FP=3,\) and \(FN=4\). Putting these values into the formulae for precision and recall, we obtain:
\[\begin{aligned} \operatorname{ Precision } &=\frac{TP}{TP + FP} \\ &=\frac{5}{5+3} \\ &=0.625 \\ \operatorname{Recall} &=\frac{TP}{TP + FN} \\ &=\frac{5}{5+4} \\ &=0.56 \end{aligned}\] 
Next, the relevant values for calculating sensitivity and specificity are \(TP=5, FP=3,\) and \(TN=8\). Note that sensitivity comes out as the same value as recall, as expected:
\[\begin{aligned} \operatorname { Sensitivity } &=\frac{TP}{TP + FN} \\ &=\frac{5}{5+4} \\ &=0.56 \end{aligned}\] whereas specificity gives:
Applications in Information Retrieval
 Precision and recall are best known for their use in evaluating search engines and other information retrieval systems.
 Search engines must index large numbers of documents, and display a small number of relevant results to a user on demand. It is important for the user experience to ensure that both all relevant results are identified, and that as few as possible irrelevant documents are displayed to the user. For this reason, precision and recall are the natural choice for quantifying the performance of a search engine, with some small modifications.
 Over 90% of users do not look past the first page of results. This means that the results on the second and third pages are not very relevant for evaluating a search engine in practice. For this reason, rather than calculating the standard precision and recall, we often calculate the precision for the first 10 results and call this precision @ 10. This allows us to have a measure of the precision that is more relevant to the user experience, for a user who is unlikely to look past the first page. Generalizing this, the precision for the first \(k\) results is called the precision @ \(k\).
 In fact, search engine overall performance is often expressed as mean average precision, which is the average of precision @ \(k\), for a number of \(k\) values, and for a large set of search queries. This allows an evaluation of the search precision taking into account a variety of different user queries, and the possibility of users remaining on the first results page, vs. scrolling through to the subsequent results pages.
Receiver Operating Characteristic (ROC) Curve
 Suppose we have the probability prediction for each class in a multiclass classification problem, and as the next step, we need to calibrate the threshold on how to interpret the probabilities. Do we predict a positive outcome if the probability prediction is greater than 0.5 or 0.3? The Receiver Operating Characteristic (ROC) curve ROC helps answer this question.
 Adjusting threshold values like this enables us to improve either precision or recall at the expense of the other. For this reason, it is useful to have a clear view of how the False Positive Rate and True Positive Rate vary together.
 The ROC curve shows the variation of the error rates for all values of the manuallydefined threshold. The curve is a plot of the False Positive Rate (also called the False Acceptance Rate) on the Xaxis versus the True Positive Rate on the Yaxis for a number of different candidate threshold values between 0.0 and 1.0. A data analyst may plot the ROC curve and choose a threshold that gives a desirable balance between the false positives and false negatives.
 False Positive Rate (also called the False Acceptance Rate) on the Xaxis: the False Positive Rate is also referred to as the inverted specificity where specificity is the total number of true negatives divided by the sum of the number of true negatives and false positives.
 True Positive Rate on the Yaxis: the True Positive Rate is calculated as the number of true positives divided by the sum of the number of true positives and the number of false negatives. It describes how good the model is at predicting the positive class when the actual outcome is positive.
 Note that both the False Positive Rate and the True Positive Rate are calculated for different probability thresholds.
 As another example, if a search engine assigns a score to all candidate documents that it has retrieved, we can set the search engine to display all documents with a score greater than 10, or 11, or 12. The freedom to set this threshold value generates a smooth curve as below. The figure below shows a ROC curve for a binary classifier with AUC = 0.93. The orange line shows the model’s false positive and false negative rates, and the dotted blue line is the baseline of a random classifier with zero predictive power, achieving AUC = 0.5.
 Note that another way to obtain FPR and TPR is through TNR and FNR respectively, as follows:
 The \(y = x\) line in the ROC curve signifies the performance of a random classifier (image credit). An ROC curve for an ideal/perfect classifier would nudge towards the topleft (since higher TPR and lower FPR is desirable) to yield AUC \(\approx\) 1:
Equal Error Rate (EER)
 The equal error rate (EER) in ROC curves is the intersection of the \(y + x = 1\) line with the DET curve (image credit):
Area under the ROC Curve (AUROC)
 The area under the ROC curve (AUROC) is a good metric for measuring the classifier’s performance. This value is normally between 0.5 (for a bad classifier) and 1.0 (a perfect classifier). The better the classifier, the higher the AUC and the closer the ROC curve will be to the top left corner.
Detection error tradeoff (DET) curve
 A detection error tradeoff (DET) curve is a graphical plot of error rates for binary classification systems, plotting the false rejection rate (FRR) vs. false acceptance rate (FAR) for different probability thresholds.
 The X and Yaxes are scaled nonlinearly by their standard normal deviates (or just by logarithmic transformation), yielding tradeoff curves that are more linear than ROC curves, and use most of the image area to highlight the differences of importance in the critical operating region.
Comparing ROC and DET curves

Let’s compare receiver operating characteristic (ROC) and detection error tradeoff (DET) curves for different classification algorithms for the same classification task.

DET curves are commonly plotted in normal deviate scale. To achieve this the DET display transforms the error rates as returned by sklearn’s det_curve and the axis scale using
scipy.stats.norm
. 
The point of this example is to demonstrate two properties of DET curves, namely:

It might be easier to visually assess the overall performance of different classification algorithms using DET curves over ROC curves. Due to the linear scale used for plotting ROC curves, different classifiers usually only differ in the top left corner of the graph and appear similar for a large part of the plot. On the other hand, because DET curves represent straight lines in normal deviate scale. As such, they tend to be distinguishable as a whole and the area of interest spans a large part of the plot.

DET curves give the user direct feedback of the detection error tradeoff to aid in operating point analysis. The user can deduct directly from the DETcurve plot at which rate falsenegative error rate will improve when willing to accept an increase in falsepositive error rate (or viceversa).


The plots in this example compare the ROC curve on the left with the corresponding DET curve on the right. There is no particular reason why these classifiers have been chosen for the example plot over other classifiers available in scikitlearn.
 To generate DET curves using scikitlearn:
import numpy as np
from sklearn.metrics import det_curve
y_true = np.array([0, 0, 1, 1])
y_scores = np.array([0.1, 0.4, 0.35, 0.8])
fpr, fnr, thresholds = det_curve(y_true, y_scores)
# fpr: array([0.5, 0.5, 0. ])
# fnr: array([0. , 0.5, 0.5])
# thresholds: array([0.35, 0.4 , 0.8 ])

Note the formulae to obtain FAR (FPR) and FRR (FNR):
\[FAR = FPR = \frac{FP}{\text{number of negatives}} = \frac{FP}{FP + TN} \\ FRR = FNR = \frac{FP}{\text{number of positives}} = \frac{FN}{FN + TP}\] where, FP: False positive; FN: False Negative; TN: True Negative; TP: True Positive

Another way to obtain FAR and FRR is through TNR and TPR respectively, as follows:
Equal Error Rate (EER)
 The equal error rate (EER) in DET curves is the intersection of the \(y = x\) line with the DET curve:
Example Walkthrough with Code
Dataset
 Let’s first generate a 2 class imbalanced dataset
X, y = make_classification(n_samples=10000, n_classes=2, weights=[0.95,0.05], random_state=42)
trainX, testX, trainy, testy = train_test_split(X, y, test_size=0.2, random_state=2)
Train a model for classification
model = LogisticRegression()
model.fit(trainX, trainy)
predictions = model.predict(testX)
Comparing Accuracy vs. PrecisionRecall with imbalanced data
accuracy = accuracy_score(testy, predictions)
print('Accuracy: %.3f' % accuracy)
 which outputs:
Accuracy: 0.957
print(classification_report(testy, predictions))

which outputs:
precision recall f1score support 0 0.96 0.99 0.98 1884 1 0.73 0.41 0.53 116 avg / total 0.95 0.96 0.95 2000
ROC Curve vs. PrecisionRecall Curve with imbalanced data
probs = model.predict_proba(testX)
probs = probs[:, 1]
fpr, tpr, thresholds = roc_curve(testy, probs)
pyplot.plot([0, 1], [0, 1], linestyle='')
pyplot.plot(fpr, tpr, marker='.')
pyplot.show()
auc_score = roc_auc_score(testy, probs)
print('AUC: %.3f' % auc_score)

which outputs:
AUC: 0.920
PrecisionRecall curve
precision, recall, thresholds = precision_recall_curve(testy, probs)
auc_score = auc(recall, precision)
pyplot.plot([0, 1], [0.5, 0.5], linestyle='')
pyplot.plot(recall, precision, marker='.')
pyplot.show()
print('AUC: %.3f' % auc_score)

which outputs:
AUC: 0.577
Evaluation Metrics for the Regression Problem
 Regression is a type of machine learning problem which helps in finding the relationship between independent and dependent variable.
 Examples include predicting continuous values such as price, rating, etc.
Mean Absolute Error (MAE)
 MAE is a very simple metric which calculates the absolute difference between actual and predicted values.
 Let’s take an example you have input data and output data and use Linear Regression, which draws a bestfit line. To find the MAE for your model, calculate the difference between the actual value and predicted value which yields the absolute error for the current sample. Repeating this for the entire dataset yields the MAE for the model. In other words, sum all the individual errors and divide them by the total number of observations.

Note that our aim is to minimize MAE because this is a loss function.
\[\text { MAE }=\frac{1}{N} \sumy\hat{y}\] where, \(N\) are the total number of data points, \(y\) is the actual output, \(\hat{y}\) is the predicted output and \(\y\hat{y}\\) is the absolute value of the residual.
from sklearn.metrics import mean_absolute_error
print("MAE:", mean_absolute_error(y_test, y_pred))
Pros of MAE
 MAE follows the same units as the output variable so it is easy to interpret.
 MAE is robust to outliers.
Cons of MAE
 The graph of MAE is not differentiable so we have to apply various optimizers like gradient descent which can be differentiable.
Mean Squared Error (MSE)
 To overcome the disadvantage of MAE, next metric came as MSE.
 MSE is widely used and differs from MAE just a little bit where MAE utilizes the absolute difference while MSE is based on the squared difference. MSE can thus be obtained by calculating the squared difference between the actual and predicted value.

Note that MSE calculates the squared distance between the actual and predicted values because it disregard the sign of the error and focuses on only the magnitude. This has the effect of avoiding the cancellation of similar positive and negative terms.
\[MSE=\frac{1}{n} \sum(y\widehat{y})^{2}\] where, \((y\widehat{y})^{2}\) is the square of the difference between the actual and predicted value.
 To obtain RMSE, we can use the NumPy square root function over MSE:
from sklearn.metrics import mean_squared_error
print("MSE:", mean_squared_error(y_test, y_pred))
Pros of MSE
 The graph of MSE is differentiable, so you can easily use it as a loss function in deep learning.
Cons of MSE
 The value you obtain after calculating MSE is a squared unit of the output. For example, if the output units are meters, then the MSE is in metersquared. This makes interpretation of the loss value difficult.
 The higher the error, more the loss. As such, if you have outliers in the dataset, MSE penalizes the outliers the most since the calculated MSE is larger.
Root Mean Squared Error (RMSE)
 As the name suggests, RMSE is the square root of MSE. RMSE is probably the most common evaluation metric when working with deep learning techniques.
import numpy as np
print("RMSE:", np.sqrt(mean_squared_error(y_test, y_pred)))
Pros of RMSE
 The RMSE uses the same units as the output variable which makes interpretation of loss easy.
Cons of RMSE
 Not that robust to outliers compared to MAE.
Root Mean Squared Log Error (RMSLE)
 Taking the log of the RMSE metric slows down the scale of error. The metric is very helpful when you are developing a generative model that is calling the inputs. In that case, the output will vary on a large scale.
 To obtain RMSLE, we can use the NumPy log function over RMSE:
print("RMSE",np.log(np.sqrt(mean_squared_error(y_test,y_pred))))
RSquared
 The Rsquared (\(R^2\)) score is a metric that measures performance by comparing your model with a baseline model.
 \(R^2\) calculates the difference between a regression line and a mean line.

\(R^2\) is also known as “coefficient of determination” or sometimes also known as goodness of fit.
\[R2 Squared=1\frac{\text { SS_r }}{\text { SS_m }}\] where, \(SS_r\) = squared sum error of the regression line; \(SS_m\) is the squared sum error of the mean line.
from sklearn.metrics import r2_score
r2 = r2_score(y_test,y_pred)
print(r2)
 The most common interpretation of rsquared is how well the regression model fits the observed data. For example, an rsquared of 60% reveals that 60% of the data fits the regression model. Generally, a higher rsquared indicates a better fit for the model.
 However, it is not always the case that a high rsquared is good for the regression model. The quality of the statistical measure depends on many factors, such as the nature of the variables employed in the model, the units of measure of the variables, and the applied data transformation. Thus, sometimes, a high rsquared can indicate the problems with the regression model.
 A low rsquared figure is generally a bad sign for predictive models. However, in some cases, a good model may show a small value.
Adjusted RSquared
 The disadvantage of the R2 score is that when adding new features to the data, the R2 score either increases or remains constant, but it never decreases because it assumes that adding more data causes the variance of data to increase.
 But the problem is when we add an irrelevant feature in the dataset, R2 sometimes starts increasing, which is incorrect.

The Adjusted R Squared metric fixes this problem.
\[\mathrm{R}_{\mathrm{a}}^{2}=1\left[\left(\frac{\mathrm{n}1}{\mathrm{n}\mathrm{k}1}\right) \times\left(1\mathrm{R}^{2}\right)\right]\] where \(n\) is the number of observations; \(k\) is the number of independent variables; \(R_a^2\) = adjusted \(R^2\).
n=40
k=2
adj_r2_score = 1  ((1r2)*(n1)/(nk1))
print(adj_r2_score)
 As \(k\) increases by adding some features, the denominator will decrease, \(n1\) will remain constant. R2 score will remain constant or will increase slightly so the complete answer will increase. When we subtract this from one then the resultant score will decrease – which is what we want when adding irrelevant features to the dataset.
 If we add a relevant feature then the R2 score will increase and \((1R2)\) will decrease heavily and the denominator will also decrease so the complete term decreases, and on subtracting from one the score increases.
Object Detection: IoU, AP, and mAP
 In object detection, two primary metrics are used: intersection over union (IoU) and mean average precision (mAP). Let’s walk through a small example.
Intersection over Union (IoU)
 Object detection involves finding objects, classifying them, and localizing them by drawing bounding boxes around them. IoU is an intuitive metric that measures the goodness of fit of a bounding box (figure credit to J. Hui’s excellent post):
 The higher the IoU, the better the fit. IoU is a great metric since it works well for any size and shape of object. This perobject metric, along with precision and recall, form the basis for the full object detection metric, mean average precision (mAP).
Average Precision (AP): Area Under the Curve (AUC)

Object detectors create multiple predictions: each image can have multiple predicted objects, and there are many images to run inference on. Each predicted object has a confidence assigned with it: this is how confident the detector is in its prediction.

We can choose different confidence thresholds to use, to decide which predictions to accept from the detector. For instance, if we set the threshold to 0.7, then any predictions with confidence greater than 0.7 are accepted, and the low confidence predictions are discarded. Since there are so many different thresholds to choose, how do we summarize the performance of the detector?

The answer uses a precisionrecall curve. At each confidence threshold, we measure the precision and recall of the detector, giving us one data point. If we connect these points together, one for each threshold, we get a precision recall curve like the following:
 The better the model, higher the precision and recall at its points: this pushes the boundary of the curve (the dark line) towards the top and right. We can summarize the performance of the model with one metric, by taking the area under the curve (shown in blue). This gives us a number between 0 and 1, where higher is better. This metric is commonly known as average precision (AP).
Mean Average Precision (mAP)

Object detection is a complex task: we want to accurately detect all the objects in an image, draw accurate bounding boxes around each one, and accurately predict each object’s class. We can actually encapsulate all of this into one metric: mean average precision (mAP).

To start, let’s compute AP for a single image and class. Imagine our network predicts 10 objects of some class in an image: each prediction is a single bounding box, predicted class, and predicted confidence (how confident the network is in its prediction).

We start with IoU to decide if each prediction is correct or not. For a ground truth object and nearby prediction, if,
 the predicted class matches the actual class, and
 the IoU is greater than a threshold,
 … we say that the network got that prediction right (true positive). Otherwise, the prediction is a false positive.

We can now sort our predictions by their confidence, descending, resulting in the following table. Table of predictions, from most confident to least confident. Cumulative recision and recall shown on the right:
 For each confidence level (starting from largest to smallest), we compute the precision and recall up to that point. If we graph this, we get the raw precisionrecall curve for this image and class:
 Notice how our precisionrecall curve is jagged: this is due to some predictions being correct (increasing recall) and others being incorrect (decreasing precision). We smooth out the kinks in this graph to produce our network’s final PR curve for this image and class. The smoothed precisionrecall curve used to calculate average precision (area under the curve):

The average precision (AP) for this image and class is the area under this smoothed curve.

To compute the mean average precision over the whole dataset, we average the AP for each image and class, giving us one single metric of our network’s performance on classification! This is the metric that is used for common object detection benchmarks such as Pascal VOC and COCO.
Evaluation Metrics for NLP/Speech Models
 Classification
 Speech Recognition
 Generative language models
 Machine Translation/Captioning
 Text Summarization
 Multitask learning
 Manual evaluation by humans for fluency, grammar, comparative ranking, etc.
Evaluation Metrics for Compression Models
Evaluation Metrics for GANbased Models
 Inception Score
 Wasserstein distance
 Fréchet Inception Distance
 Traditional metrics such as precision, recall, and \(F_1\) score
Fréchet Inception Distance
 This metric compares the statistics of the generated samples and real samples. It models both distributions as multivariate Gaussian. Thus, these two distributions can be compactly represented by their mean \(\mu\) and covariance matrix \(\Sigma\) exclusively. That is:

These two distributions are estimated with 2048dimensional activations of the Inceptionv3 pool3 layer for real and generated samples respectively.

Finally, the FID between the real image distribution (\(Xr\)) and the generated image distribution (\(Xg\)) is computed as:
 Therefore, lower FID corresponds to more similar real and generated samples as measured by the distance between their activation distributions.
Evaluation Metrics for recommender systems
 In A survey on news recommender systems (2020), Raza and Ding offer a summary of the definitions for a range of evaluation metrics in the conext of recommender systems:
Metric  Description  Type 

Precision  The proportion of relevant recommended items over total recommended items.  Accuracy 
Recall  The proportion of relevant recommended items over total relevant items.  Accuracy 
F1score  Weighted average of the precision and recall.  Accuracy 
Customer Satisfaction Index  The satisfaction degree of a user on the recommendations (Xia et al. 2010).  Beyondaccuracy 
Mean Reciprocal Rank (MRR)  The multiplicative inverse of the rank of the first correct item.  Ranking accuracy 
Mean Average Precision (MAP)  The average precisions across all relevant queries.  Ranking accuracy 
\(\overline{Rank}\)  The percentileranking of article within the ordered list of all articles.  Ranking accuracy 
Cumulative rating  The total relevance of all documents at or above each rank position in the top \(k\).  Ranking accuracy 
Success @ \(k\)  A current news item that is in sequence and in a set of recommended news items.  Ranking accuracy 
Personalized @ \(k\)  A current news item that is in a given sequence and in a set of recommended news items without popular items (Garcin et al. 2013).  Personalization accuracy 
Novelty @ \(k\)  The ratio of unseen and recommended items over the recommended items.  Novelty, beyondaccuracy 
Diversity  The degree of how much dissimilar recommended items are for a user.  Diversity, beyondaccuracy 
Binary Hit rate  The number of hits in an nsized list of ranked items over the number of users for whom the recommendations are produced.  Ranking accuracy 
Logloss  To measure the performance of a classification model where the prediction input is a probability value between 0 and 1.  Accuracy 
Average Reciprocal Hitrate  Each hit is inversely weighted relative to its position in topN recommendations.  Ranking accuracy 
Rootmeansquare error (RMSE)  Difference between the predicted and the actual rating.  Accuracy 
Clickthrough rate (CTR)  The likelihood of a news item that will be clicked.  Accuracy 
Discounted Cumulative Gain (DCG)  The gain of an item according to its position in the result list of a recommender.  Ranking accuracy 
Area under curve (AUC)  A ROC curve plots recall (True Positive Rate) against fallout (False Positive Rate).  Accuracy 
Saliency  To evaluate if a news entity is relevant for a text document (Cucchiarelli et al. 2018).  Beyondaccuracy 
FutureImpact  To evaluate how much user attention (views or shares) each news story may receive in the future and is measured between recency and relevancy (Chakraborty et al. 2019).  Beyondaccuracy 
Further reading
Here are some (optional) links you may find interesting for further reading:
 An empirical study on evaluation metrics of generative adversarial networks (paper)
 How to measure GAN performance? (blog post)
 Are GANs Created Equal? A LargeScale Study (paper)
 Pros and Cons of GAN Evaluation Measures (paper)
 How to Evaluate GANs (blog post)
 Evaluation Metrics for Language Modeling (article)
 Evaluation Metrics for Language Models (paper)
 Analytics Vidhya: Know The Best Evaluation Metrics for Your Regression Model
 The Ultimate Performance Metric in NLP
References
 Speech and Language Processing (2019) by Jurafsky and Martin.
 A Probabilistic Interpretation of Precision, Recall and FScore, with Implication for Evaluation (2005) by Goutte and Gaussier.
 Information Retrieval (2nd ed.) (1979) by Van Rijsbergen.
 Machine literature searching VIII. Operational criteria for designing information retrieval systems (1955) by Kent et al.
 CS229 Machine Learning (Tips and Tricks Cheatsheet)
 ROC Curves and PrecisionRecall Curves for Classification
 Wikipedia: Precision and Recall
 Wikipedia: Fscore
 Wikipedia: Sensitivity and Specificity
 The PrecisionRecall Plot Is More Informative than the ROC Plot when Evaluating Binary Classifiers on Imbalanced Datasets
 F1 Score vs ROC AUC vs Accuracy vs PR AUC: Which Evaluation Metric Should You Choose?
 sklearn.metrics.det_curve
 Wikipedia: Detection error tradeoff
 Are FAR and FRR the same as FPR and FNR, respectively?
Citation
If you found our work useful, please cite it as:
@article{Chadha2020DistilledEvaluationMetrics,
title = {Evaluation Metrics, ROCCurves and Imbalanced Datasets},
author = {Chadha, Aman},
journal = {Distilled AI},
year = {2020},
note = {\url{https://aman.ai}}
}