Latent semantic analysis persona clustering

Description

Cluster participants into personas by the semantic similarity of their free-text answers, computed with latent semantic analysis, then check agreement with expert groupings.

Participants answer semi-structured questions about goals, needs and frustrations. Their texts are compared pairwise with latent semantic analysis, giving a cosine similarity matrix between participants, which is clustered with hierarchical agglomerative clustering; clusters become personas. Miaskiewicz et al. reported agreement with experts above kappa 0.60, but Brickey, Walczak and Burgess found only fair agreement (kappa 0.31) and lopsided clusters (12, 5 and 1) with 18 participants. Results depend on the reference corpus behind the semantic space, and the method needs specialist tools. LSA has since been superseded by word and sentence embeddings, so the approach survives mainly as a precursor to embedding-based clustering of qualitative text.

Attributes

Loss type
Extinct
Data
Mixed
Effort
Days

Steps

  1. Collect free-text answers to the same semi-structured questions about goals, needs and frustrations.
  2. Choose or build a semantic space, preferably from a domain corpus: weight the term counts (for example by tf-idf, so that shared rare words count more) and reduce them by singular value decomposition.
  3. Compute LSA cosine similarities between every pair of participants' texts.
  4. Run hierarchical agglomerative clustering on the similarity matrix and choose the number of clusters.
  5. Check agreement with an expert clustering using Cohen's kappa, aiming for at least 0.6.
  6. Turn acceptable clusters into personas, reviewing unbalanced or singleton clusters by hand.

Sources

  • A Comparative Analysis of Persona Clustering Methods (Brickey et al., 2010)
  • Computer Vision: A Modern Approach (Forsyth & Ponce, 2011)

Also known as

  • LSA persona clustering
  • Semantic similarity clustering of interview text

Categories

  • Analytics & Behaviour
  • Computational & Machine Learning