Lecture 6 · Interactive explanation

Count predictions; then discover groups

Interactive labLecture 6

These are two different tasks. Classification has supplied labels to check against. Clustering forms groups without being given the answers.

1. A confusion matrix

Take injured as positive. Enter test-set counts. Precision starts with positive predictions; recall starts with actual positives.

True positives

False positives

False negatives

True negatives

Accuracy

Precision

Recall

If a denominator is zero, the corresponding metric is undefined here. Software sometimes assigns a convention instead; check which convention is used.

2. k-means, one iteration at a time

These are invented points in two features measured on a comparable scale. Coloured dots are assignments; large markers are centroids.

Number of clusters k

Show PCA axes on the same points

k is chosen by you here, not proved by the algorithm. Lower within-cluster error as k grows does not prove more meaningful groups. PCA is another unsupervised operation: its axes retain variance; they do not supply class labels.