T-SNE

Description

A non-linear method that maps high-dimensional user data to two dimensions while preserving local neighbourhoods, for visual exploration of behavioural groupings.

t-SNE converts pairwise distances into probabilities of being neighbours in high-dimensional space and finds a 2D layout whose neighbour probabilities match, using a heavy-tailed t-distribution to avoid crowding. It reveals local cluster structure that PCA can flatten. But its maps are easily misread: distances between clusters, cluster sizes and densities carry little meaning, results vary with the perplexity parameter and random seed, and it can produce apparent clusters from noise. Rigour means using it for exploration only, running several perplexities (e.g. 5, 30, 50) and seeds, initialising with PCA, and confirming any visual grouping with clustering and validation on the original features.

Attributes

Loss type
Active
Effort
Hours

Steps

  1. Standardise features and reduce to around 30-50 dimensions with PCA to remove noise and speed computation.
  2. Run t-SNE with perplexity values of 5, 30 and 50, PCA initialisation and at least 1,000 iterations.
  3. Repeat with several random seeds and compare whether the same groupings recur.
  4. Colour maps by known variables (outcomes, segments, platforms) to interpret regions.
  5. Do not interpret inter-cluster distances or cluster sizes; confirm any apparent groups with clustering and validation on original data.

Sources

  • Pattern Recognition and Machine Learning (Christopher Bishop)
  • Unsupervised Learning in Space and Time (Marc Leordeanu)
  • van der Maaten & Hinton (2008) Visualizing data using t-SNE
  • Wattenberg, Viégas & Johnson (2016) How to use t-SNE effectively
  • Pattern Recognition & ML (Bishop)
  • Unsupervised Learning in Time & Space
  • The Inquiry Series, Ch 52 — Dimensionality Reduction — The Geometry of Behavioral Space

Also known as

  • t-distributed stochastic neighbour embedding
  • t-SNE

Categories

  • Analytics & Behaviour
  • Computational & Machine Learning