Interactive clustering canvas
Two-dimensional feature projection
Cluster sizes
Cluster profile radar
Similarity matrix
Current animal dataset
| Animal | Weight | Speed | Habitat | Diet | Activity | Social | Cluster |
|---|
Cluster evaluation
Quality history
Selected animals
Shared and different traits
How clustering works
Clustering is unsupervised learning because the dataset has no predefined target labels. The algorithm searches for animals that are similar across the selected physical, habitat, diet, and behavioural features.
K-means
Assigns animals to the nearest centroid, then repeatedly recalculates each centroid. It works best with compact, similarly sized groups.
K-medoids
Uses real animals as group representatives. It is often more robust when extreme animals distort average centroids.
Hierarchical clustering
Starts with individual animals and merges the closest groups. It reveals nested relationships and supports dendrogram-style thinking.
DBSCAN
Builds dense neighbourhoods and marks isolated animals as outliers. It can find irregular groups without choosing a fixed cluster count.
Why preprocessing matters
Weight can reach thousands while social behaviour uses a ten-point scale. Normalisation prevents large numerical ranges from dominating every distance calculation. Categorical encoding converts habitat, diet, cover, and activity into machine-readable dimensions.