Skip to content

Specificity (TNR)#

Specificity, also known as true negative rate (TNR), measures the proportion of negative ground truths that a model correctly predicts, ranging from 0 to 1. A high specificity indicates that the model is effective at correctly identifying negative cases, where a low specificity suggests that the model is misclassifying many negative cases as positive.

As shown in this diagram, specificty is the fraction of all negative ground truths that are correctly predicted:

\[\text{Specificity} = \frac{\text{TN}}{\text{TN} + \text{FP}}\]

In the above formula, \(\text{TN}\) is the number of true negative inferences and \(\text{FP}\) is the number of false positive inferences.

Guide: True Negative / False Positive

Read the TP / FP / FN / TN guide if you're not familiar with "TN" and "FP" terminology.

Specificity Image Specificity Image

Specificity is often used in conjunction with sensitivity (recall), also known true positive rate (TPR). By measuring both specificity and sensitivity, a more complete picture of a model's performance can be drawn.

Implementation Details#

Specificity is used to evaluate the performance of a classification model, particularly in tasks like binary classification, where the goal is to classify data into one of two possible classes. It is especially useful when the objective is to measure the model's ability to correctly identify the negative class instances. For example, in medical diagnostics, a high specificity is essential to avoid unnecessary treatments or alarm.

Here is how specificity is calculated:

\[ \text{Specificity} = \frac {\text{# True Negatives}} {\text{# True Negatives} + \text{# False Positives}} \]


Perfect model inferences, where every negative ground truth is recalled by an inference:

Metric Value
TN 20
FP 0
\[ \begin{align} \text{Specificity} &= \frac{20}{20 + 0} \\[1em] &= 1.0 \end{align} \]

Partially correct inferences, where some negative ground truths are correctly recalled (TN) and others are missed (FP):

Metric Value
TN 85
FP 15
\[ \begin{align} \text{Specificity} &= \frac{85}{85 + 15} \\[1em] &= 0.85 \end{align} \]

Zero correct inferences — no negative ground truths are recalled:

Metric Value
TN 0
FP 20
\[ \begin{align} \text{Specificity} &= \frac{0}{0 + 20} \\[1em] &= 0.0 \end{align} \]

Multiple Classes#

So far, we have only looked at binary classification cases, but in multiclass or multi-label cases, specificity is computed per class. In the TP / FP / FN / TN guide, we went over multiple-class cases and how these metrics are computed. Once you have these four metrics computed per class, you can compute specificity for each class by treating each as a single-class problem.

Aggregating Per-class Metrics#

If you are looking for a single specificity score that summarizes model performance across all classes, there are different ways to aggregate per-class specificity scores: macro, micro, and weighted. Read more about these methods in the Averaging Methods guide.

Limitations and Biases#

While specificity is a valuable metric for evaluating the performance of classification models, it does have limitations and potential biases that should be considered when interpreting results:

  1. Sensitivity to Class Imbalance: Specificity is sensitive to class imbalance in dataset. If one class significantly outnumbers the other, a high specificity can be achieved simply by predicting the majority class most of the time. This can lead to a misleadingly high specificity score while neglecting the model's ability to correctly classify the minority class.

  2. Ignoring False Negatives: Specificity focuses exclusively on the true negatives (correctly classified negative cases) and false positives (negative cases incorrectly classified as positive), but it doesn't account for false negatives (positive cases incorrectly classified as negative). Ignoring false negatives can be problematic in applications where missing positive cases is costly or has severe consequences.

  3. Incomplete Context: Specificity alone does not provide a complete picture of a model's performance. It is often used in conjunction with other metrics like sensitivity (recall), precision, and F1-score to provide a more comprehensive assessment. Depending solely on specificity might hide issues related to other aspects of classification, such as the models' ability to identify true positives.

  4. Threshold Dependence: Specificity is a binary metric that doesn't take into account the probability or confidence levels associated with predictions. Models with different probability distributions might achieve the same specificity score, but their operational characteristics can vary significantly. To address this limitation, consider using threshold-independent metrics like the area under the receiver operating characteristic curve (AUC-ROC) which can provide a more comprehensive understanding of model performance.