Description
Fit a tree to pairwise similarity data in which item distance equals path length, revealing distinct groups and items typical of many others.
Unlike hierarchical clustering, which forces all items to the same depth (an ultrametric), an additive tree lets branch lengths vary, so distance between two items is the length of the path joining them. Long internal branches show distinct clusters; short leaf branches mark items close to many others. Sattath and Tversky build the tree from quadruples: for every four items they find which pairs are neighbours, merge the pair that wins most often, and repeat, then estimate arc lengths by least squares. On animals, letters and an occupation sort by 154 people, trees matched or beat two- and three-dimensional spatial solutions on fit and were easier to read; structure survived noise of a third of each distance. Skewness of the distances helps choose between a tree and a spatial map, and where the root is placed is an interpretive choice.
Attributes
- Discipline
- Statistics & Modelling
- Loss type
- Unmigrated
- Stage
- Definition, Design
- Data
- Quantitative
- Effort
- Days
Steps
- Collect a symmetric proximity matrix: similarity ratings, card-sort co-occurrence counts or confusion data.
- Check whether the data look tree-like or space-like using the skewness of distances (tree distances bunch at the top of the range, spatial distances at the bottom).
- Fit the additive tree, estimate arc lengths and set negative arcs to zero.
- Report stress and the correlation with the data, and compare with hierarchical clustering and a two-dimensional map.
- Try more than one root and label clusters only where other evidence supports the grouping.
Sources
- Preference, Belief, and Similarity (Tversky, 2004)
Also known as
- ADDTREE
- Additive tree clustering
- Path-length tree fitting
- ADDTREE scaling
Categories
- Analytics & Behaviour
- Statistics & Modelling