Fuzzy C-Means — membership instead of a hard label
Fuzzy C-Means keeps a membership vector for every row. Memberships are positive and sum to 1. The update weights each row's pull on a center by membership raised to m. At m=2 a row that sits between two centers contributes to both. Hardening the memberships (taking the largest) produces the label plotted on the chart; the memberships themselves are the result worth keeping when the boundary is genuinely soft.
On scaled Iris, c=3 and m=2, the hardened labels scored silhouette 0.458 and adjusted Rand 0.630 against species, almost the same partition as K-Means. The extra information is the membership of the rows near the versicolor/virginica boundary, which K-Means would have snapped to one centroid with no record of the doubt.
As m approaches 1 the memberships become hard and the algorithm collapses toward K-Means. As m grows, every membership drifts toward 1/c and the centers slide together.
Settings that change this fit
- c
- How many fuzzy centers are updated. Same role as k, but every row keeps a weight for each center.
- m
- Fuzzifier. 2 is the usual setting. Values near 1 harden the memberships; values much above 2 wash them out.