Understanding Model Performance: Precision, Recall, and F1 Score Explained

1 Model Performance

Classification is the process of assigning data points to classes. In machine learning we train a model and then measure its performance using loss functions, but how do we actually evaluate it? Accuracy alone is insufficient.

For example, if a dataset contains over 90% of class 0 and the rest class 1, a naive model that always predicts 0 would achieve over 90% accuracy, showing that accuracy does not reveal the whole picture.

2 Confusion Matrix

A better way to assess a classifier is to examine its confusion matrix, which counts how many instances of class A are predicted as class B.

Each row of the matrix corresponds to an actual class, each column to a predicted class. While the matrix provides rich information, we often summarize it with specific metrics.

Precision Precision = TP / (TP + FP). It measures the proportion of positive predictions that are correct. A perfect precision of 100% can be achieved by predicting only one positive instance correctly, but this may cause the classifier to ignore other instances.

Recall Recall = TP / (TP + FN). It measures the proportion of actual positives that are correctly identified.

Assume a classifier has a precision of 72.9% and a recall of 75.6%. How can we combine these two criteria?

We can use the F1 score , the harmonic mean of precision and recall.

The F1 score favors classifiers with balanced precision and recall. In some cases you may prioritize precision (e.g., filtering unsafe videos for children) while in others recall is more important (e.g., detecting shoplifters where missing a true case is costly).

Improving precision typically reduces recall and vice‑versa; they cannot be simultaneously maximized.

References

https://www.geeksforgeeks.org/confusion-matrix-machine-learning/?ref=lbp