Choosing a Resemblance Coefficient for Clustering

Description

Select the similarity or distance measure for clustering by deciding which profile differences should count, then confirm the choice on small toy matrices whose right answer you can judge.

Distance and similarity coefficients differ in what they ignore. Euclidean distance ignores nothing; Penrose's shape difference ignores additive shifts; cosine ignores proportional shifts; Pearson correlation ignores both but treats mirror images as opposites; Canberra weights attributes equally; Bray-Curtis is dominated by large values. For binary data the decision is whether shared absences (0-0 matches) are evidence of similarity: Jaccard and Sorensen exclude them, simple matching includes them. Romesburg's procedure is to decide from the research goal whether level and size differences matter, shortlist coefficients with the right sensitivities, build small made-up matrices representing plausible outcomes, judge by intuition which objects should group, and keep the coefficient that agrees. He warns that several coefficients are monotonically related, so obtaining the same tree from them is no confirmation.

Attributes

Loss type
Latent
Effort
Hours

Steps

  1. State whether differences in overall level or size between profiles are meaningful for the research goal.
  2. For binary data, decide whether shared absence of a feature should count as similarity.
  3. Shortlist coefficients with the required sensitivities (for example cosine to ignore proportional shifts, Jaccard to exclude 0-0 matches).
  4. Drop attributes that never vary and use Jaccard-type coefficients for dummy-coded multistate attributes.
  5. Build several small toy matrices of plausible outcomes and decide by intuition which objects belong together.
  6. Compute each shortlisted coefficient on the toy matrices and keep the one whose grouping matches; use average Euclidean distance when values are missing.

Sources

  • Cluster Analysis for Researchers (Romesburg, 1984)

Also known as

  • Choosing a resemblance coefficient by sensitivity to profile shifts and toy-matrix calibration
  • Resemblance coefficient choice
  • Size displacement sensitivity
  • 0-0 match decision
  • Euclidean, cosine, correlation, Canberra, Bray-Curtis, Jaccard, simple matching

Categories

  • Analytics & Behaviour
  • Statistics & Modelling