Home

Classification evaluation

Table of Contents

Summary

Precision, Recall, F1

Assume we have a classifier that adds labels (classes) to each document or data point. How do we evaluate how well it is doing its job? We have three cases, considering all predicted labels as a group:

True positive (tp)
predicted labels that are true labels
False positive (fp)
predicted labels that are not true labels
False negatives (fn)
true labels that were not predicted

Since these measures depend on the number of labels (predicted and true), and that may differ depending on the document/data point or experiment, we calculate the following normalized scores:

Precision
\(tp/(tp+fp)\) — higher precision means the labels that were predicted were more often true labels
Recall
\(tp/(tp+fn)\) — higher recall means more of the true labels were predicted

We can achieve very high precision by only predicting labels for which we have very high confidence (until we nearly choose no labels at all); this would decrease recall, however. We can achieve high recall by predicting every label for every data point; this would decrease precision, however.

Best would be high precision and high recall. We can measure this by combining precision and recall into a single formula, called the "F-score" or "F1-measure":

F-score
\(2*precision*recall / (precision + recall)\)

Notice that precision and recall are treated equally in this calcuation. Technically, the F-score is the harmonic mean of precision and recall.

This graph (from Wikipedia) shows the harmonic mean of two numbers (\(x\) and \(y\), which you can interpret as \(precision\) and \(recall\)). It shows the relationship we want: if either \(x\) or \(y\) is \(0\), the F-score is \(0\). Otherwise, the F-score generally increases with increases in \(x\) and \(y\).

harmonic-mean.png

k-Fold Cross Validation

When we have one dataset with the answers (the "class" for each data point), we can split this dataset into a training and testing portion. The training portion is used to build a model of the dataset, and the testing version is used to test that model. We'll want to split the dataset multiple times at random places and then average the results. That way, we can be sure that we didn't just get lucky (or unlucky) and split the dataset into training/testing portions that are not representative of all the data.

Most common is 10-fold cross validation. This means we choose 90% of the data to be the training set, and 10% to be the testing set. We evaluate the precision/recall/etc. with this split, then choose a different 90/10 split and do it again. Because there are 10 possible splits, we do it 10 times and average 10 results. The graph below (from Selecting Representative Data Sets by Borovicka, et al.) illustrates k-fold cross validation.

k-fold-cross-validation.jpg

Intro to AI material by Joshua Eckroth is licensed under a Creative Commons Attribution-ShareAlike 3.0 Unported License. Source code for this website available at GitHub.