HyPersona semi-automated clustering algorithm selection

Description

Run every combination of clustering algorithms and parameters, screen them on internal metrics and feature significance, then have an expert pick the set that yields distinct, theory-consistent personas.

HyPersona systematises the choice of clustering algorithm for quantitative personas. Clean numeric data, a dictionary of algorithms and parameter values, and optional domain features are input. Every combination is run and scored on silhouette, Calinski-Harabasz, Davies-Bouldin and a new Average Feature Significance metric (the share of features differing significantly at p < .05 across clusters and cluster pairs), plus cluster sizes. Combinations failing lenient preset thresholds are dropped with reasons recorded. For survivors, centroids are plotted as z-profiles and early-stage personas are written automatically. A domain expert merges functionally identical solutions, reviews the rest in metric order with AFS as the main guide, and judges fit to a behavioural theory and distinctness. In the published case, 12 combinations of hierarchical clustering, k-means and NMF on 519 residents yielded three personas judged against the Protective Action Decision Model.

Attributes

Loss type
Latent
Effort
Weeks

Steps

  1. Prepare data: map ordinal answers to numbers, one-hot encode nominal ones, and impute missing values.
  2. Define key and aggregate features from domain theory, and fix the number of clusters from theory.
  3. Pilot to set lenient thresholds, for example silhouette below 0, Calinski-Harabasz below 10, Davies-Bouldin above 5, AFS below half the key features, or any cluster under 5 per cent.
  4. Expand the algorithm-parameter dictionary into all combinations and run each, logging metrics and cluster sizes.
  5. Drop combinations failing thresholds, recording why.
  6. Plot centroid z-profiles and generate early-stage personas for survivors; merge functionally identical solutions.
  7. Have a domain expert review survivors in metric order against theory and distinctness, then build full personas from the chosen set.

Sources

  • Selecting a Clustering Algorithm (HyPersona) (Ditton et al., 2022)

Also known as

  • HyPersona
  • Semi-automated clustering algorithm selection for personas
  • Persona clustering hyperparameter tuning
  • Average Feature Significance (AFS) screening

Categories

  • Analytics & Behaviour
  • Computational & Machine Learning