k-Medoids (PAM) | Interactive Workspace

Watch k-Medoids (PAM) learn

Follow each update in real time. Scrub or play to inspect the algorithm.

Step 1 of 8

02 Interactive demo

CURRENT LOSS · cost

402

Total Absolute Dissimilarity Cost per PAM iteration

WHAT'S HAPPENING

1. The Data & Extreme Outliers

We have 29 data points in natural clusters, plus 2 extreme outliers in the far corners (shown in red). If we used standard K-Means, those two outliers would pull the cluster centers far out into empty space. K-Medoids fixes this by requiring centers to be actual real data points (medoids).

\mathbf{m}_k \in \{\mathbf{x}_1, \dots, \mathbf{x}_n\} \quad (\text{Centers are real points})

INTERACTIVE EXPLORER · K-MEDOIDS (PAM)

Centroid vs. Medoid

Offset: +65px

Drag the slider to move the outlier farther away. Watch how the K-Means Centroid gets pulled out into empty space, while the K-Medoids Medoid stays locked onto real point #5.

Drag Outlier Position:+65 px

Cluster Core (+0px)Extreme Outlier (+95px)

K-Medoids: Point #5 (Zero drift)

K-Means: Average (Dragged)

Outlier: Far point

K-MEANS (AVERAGE)

Drift from Cluster:+6.5 px

The outlier drags the average out into empty space. The center becomes a meaningless synthetic point.

K-MEDOIDS (TEAM CAPTAIN)

Drift from Cluster:0.0 px (0.0)

Center is locked to real member #5. The faraway outlier has zero leverage to move the captain.