How Is a Classification Report in Deep Learning Calculated? (2026)

You trained a neural network. Accuracy looks great, 95%. You feel good. You deploy.

Then someone points out that your model is completely ignoring one class. It predicted “No Fraud” 95% of the time because 95% of your dataset was not fraud. Your model learned nothing. It just guessed the majority class.

This is exactly why accuracy alone is a trap, and why the classification report in deep learning exists: it breaks down your model’s performance class by class and tells you the truth.

This article breaks down how a classification report in deep learning is calculated, what every number means, and how to use it when evaluating neural networks.

What Is a Classification Report in Deep Learning?

A classification report in deep learning is a summary of key deep learning evaluation metrics, precision, recall, F1 score, and support, calculated for every class your model predicts. It tells you not just if your model is right, but where it’s right and where it’s quietly failing.

In Python, you generate it with one line using scikit-learn:

python

from sklearn.metrics import classification_report
print(classification_report(y_test, y_pred))

The output looks something like this:

              precision    recall  f1-score   support

     Class 0       0.91      0.99      0.95     10981
     Class 1       0.80      0.25      0.38       619

    accuracy                           0.90     11600
   macro avg       0.86      0.62      0.66     11600
weighted avg       0.90      0.90      0.89     11600

Each row represents one class. Each column measures something different. And together, they give you a complete picture of your model's true performance.

First, Understand the Confusion Matrix

Before any metric makes sense, you need to understand the confusion matrix in deep learning. It’s the foundation on which everything else is built.

For binary classification, a confusion matrix gives you four values:

• True Positive (TP): Model predicted positive. It was positive. Correct.
• True Negative (TN): Model predicted negative. It was negative. Correct.
• False Positive (FP): Model predicted positive. It was actually negative. Wrong.
• False Negative (FN): Model predicted negative. It was actually positive. Wrong.

If your model is a doctor diagnosing diseases, a False Negative means telling a sick patient they’re healthy. That’s the dangerous mistake. A False Positive means telling a healthy patient they might be sick, still bad, but probably less catastrophic.

Your classification metrics are simply different ways of counting and weighing these four outcomes.

Precision

Precision answers one question: of all the times the model predicted a class, how many times was it actually correct?

The formula is:

Precision = TP / (TP + FP)

If your spam classifier flags 100 emails as spam and 80 of them are genuinely spam, your precision is 80%.

High precision means fewer false alarms. It matters most when false positives are expensive, for example, in medical diagnosis or legal document review, you don’t want innocent cases flagged wrongly.

According to Google’s Machine Learning documentation, precision and recall together form the basis of F1, and when both are perfect (1.0), the F1 score is 1.0.

Recall

Recall (also called sensitivity) asks: of all actual positive cases, how many did the model catch?

The formula is:

Recall = TP / (TP + FN)

Using verified formulas from published deep learning research: Recall, referred to as sensitivity, is defined as the ratio of true positives to the sum of true positives and false negatives.

So if there are 200 fraud transactions in your dataset and your model catches 150 of them, recall is 75%.

Low recall means your model is missing real positives. In fraud detection, cancer screening, or network security, missing a real case can be catastrophic. You want recall as high as possible in those domains.

Here is a quick way to remember the difference:

• Precision = “Don’t cry wolf” (avoid false alarms)
• Recall = “Don’t let anything slip through” (catch everything real)

F1 Score

So which matters more, precision or recall? Often, it depends on the problem. But when you want a single number that balances both, you use the F1 score.

The F1 score is the harmonic mean of precision and recall:

F1 = 2 × (Precision × Recall) / (Precision + Recall)

Or equivalently, using raw values from the confusion matrix:

F1 = 2TP / (2TP + FP + FN)

The F1 score is the harmonic mean of precision and recall. This metric balances the importance of precision and recall, and is preferable to accuracy for class-imbalanced datasets. When precision and recall are far apart, F1 will be similar to whichever metric is worse.

The F1 score punishes imbalance between precision and recall. A model with 90% precision and 10% recall doesn’t get to average out to 50%; the F1 score will be around 18%. That’s fair. A model that catches almost nothing useful is not a good model.

Accuracy Score in Deep Learning

Accuracy is the simplest metric: total correct predictions divided by total predictions.

Accuracy = (TP + TN) / (TP + TN + FP + FN)

Accuracy calculates the number of correct predictions divided by the number of predictions.

It is useful when your classes are balanced. If you have 50% cats and 50% dogs in your dataset, accuracy tells you something real.

But on an imbalanced dataset? Accuracy becomes nearly meaningless. That fraud example from the introduction? 95% accuracy, zero learning. The classification report catches this immediately, the F1 score for the minority class would be close to zero, exposing the problem accuracy hid.

Support

Support is simply the count of actual samples per class in your test set. It shows up as the last column in the classification report.

Support refers to the number of actual occurrences of the class in the dataset. For example, a support value of 1 in a class means that there is only one observation with an actual label of that class.

Why does support matter? Because it tells you how much to trust the metrics for each class. A class with support = 3 and F1 = 1.0 is not impressive; you only had 3 examples. A class with support = 5,000 and F1 = 0.94 is something to be proud of.

Always look at support before drawing conclusions about any per-class metric.

Macro Average vs. Weighted Average

At the bottom of the classification report, you see two summary rows. This is where many people get confused.

Macro Average

The macro average simply takes the arithmetic mean of the metric across all classes, treating every class equally, regardless of how many samples it has.

The macro average is computed by averaging the unweighted mean per label.

If Class A has F1 = 0.90 and Class B has F1 = 0.30, macro F1 = 0.60. Simple. It does not care that Class A has 10,000 samples and Class B only has 50.

Use macro average when every class matters equally, for example, a medical classification system where missing a rare disease is as critical as mishandling a common one.

Weighted Average

The weighted average factors in support when computing the average. Classes with more samples contribute more to the final number.

The weighted average is computed by averaging the support-weighted mean per label.

Weighted average is especially useful for imbalanced datasets because it takes into account the weight of each class.

Use a weighted average when your classes are imbalanced, and you care more about overall model performance on the actual distribution of your data.

A Real Example: Multiclass Classification Report

Let’s say you’re building an image classifier for three animal categories: cat, dog, and fish. Your model predicts 25 test images and produces this classification report:

              precision    recall  f1-score   support

         cat       0.85      0.92      0.88        13
         dog       0.78      0.80      0.79         5
        fish       0.67      0.57      0.62         7

    accuracy                           0.80        25
   macro avg       0.77      0.76      0.76        25
weighted avg       0.79      0.80      0.80        25

The model performs best on “cat”, which makes sense, as it has the most training examples (support = 13). “Fish” is struggling with both precision and recall. The macro average drops noticeably below the weighted average, which signals that performance on smaller classes is pulling the overall score down.

This is the kind of insight you simply cannot get from looking at the accuracy score alone (80% here looks okay, but the fish class is quietly failing).

For Multiclass Problems: One-vs-Rest Approach

In multiclass classification with neural networks, the classification report in deep learning computes metrics using a One-vs-Rest (OvR) strategy.

In multi-class classification, the F1 score is calculated for each class in a One-vs-Rest approach instead of a single overall F1 score. In this approach, metrics are determined for each class separately, as if there is a different classifier for each class.

In practice, this means when computing precision and recall for “cat,” the model treats it as “cat vs. everything else.” It does this for every class, then aggregates using macro or weighted averaging.

This is why the sklearn classification report scales naturally from binary to multiclass without you needing to change anything.

How to Interpret Classification Report Results

Here is a step-by-step framework for reading classification report in deep learning output on any deep learning model:

Step 1 — Check support first. Understand how many samples exist per class before judging any metric.

Step 2 — Scan per-class F1 scores. Any class with F1 below 0.70 needs attention. Below 0.50 is a serious problem.

Step 3 — Compare precision vs. recall per class. If precision is much higher than recall, the model is conservative; it’s missing positives. If recall is higher, it’s throwing too many false alarms.

Step 4 — Look at the macro vs. weighted average gap. A large gap means your model performs very differently across classes. That usually points to a class imbalance issue.

Step 5 — Don’t celebrate accuracy alone. If overall accuracy looks good, but some F1 scores are low, the accuracy is misleading you.

Classification Report for Neural Networks

Imbalanced datasets: This is the biggest one. Neural networks on imbalanced data tend to ignore minority classes. The classification report in deep learning exposes this; accuracy hides it. When dealing with imbalanced data, the weighted average gives more importance to larger classes, ensuring the final average reflects the true distribution of data.

Confusing macro and weighted averages: Use macro when all classes matter equally (rare disease detection). Use weighted when you care about the overall performance, weighted by class frequency.

Not looking at support: A class with support = 2 and a perfect F1 means almost nothing. Don’t optimize for it.

Using accuracy as the primary metric on skewed data: As established by Google’s own ML documentation, F1 is preferable to accuracy on imbalanced datasets.

Quick Reference: The Formulas

Here is every formula from the classification report in one place:

Conclusion

The classification report in deep learning is not a nice-to-have. It is the difference between knowing your model works and thinking it works.

For any deep learning or machine learning classification task, here is what to take away:

Accuracy tells you the headline. The classification report in deep learning tells you the story. Precision, recall, and F1 score together reveal which classes your model is confident about and which ones it’s quietly ignoring. Support tells you how much trust to place in each number. And macro vs. weighted averages help you understand whether you have a class imbalance problem or a genuine performance problem.

The next time you see 95% accuracy on your neural network, run the classification report before you celebrate. Your minority class might be sending an SOS.

Noman Akram

Noman Akram is the Founder and Editor-in-Chief of TWT News. He is a technology journalist with 5+ years of experience covering artificial intelligence, AI in healthcare, blockchain, cloud computing, and cybersecurity. He built TWT News to make complex emerging technologies understandable for professionals, students, and business leaders. Based in UK (United Kingdom), his reporting covers global tech developments with a focus on real world impact.