Variational Autoencoder for Behavioural Patterns

Description

Learns a compact latent space of user behaviour from session data, enabling archetype interpolation, synthetic sequences for rare segments and counterfactual exploration.

VAEs learn a low-dimensional latent representation of high-dimensional data, then use that latent space to generate new samples. Applied to UX: fit a VAE to behavioral session data. The latent space represents 'behavioral styles'. New points in the latent space generate synthetic but realistic behavioral sequences, enabling: (1) behavioral archetype interpolation, (2) data augmentation for rare user types, (3) counterfactual behavioral analysis.

Attributes

Loss type
Unmigrated
Stage
Discovery
Effort
Weeks

Steps

  1. Represent each session as a fixed-length feature vector (action counts, timing statistics, error rates, navigation patterns).
  2. Train a VAE with encoder: session → z (latent), decoder: z → session.
  3. Visualize latent space using UMAP.
  4. Identify distinct behavioral clusters in latent space.
  5. Generate synthetic sessions at cluster boundaries to understand behavioral transitions.
  6. Use for: identifying underrepresented user types, generating training data for classifiers.

Sources

  • Deep Generative Modelling — Springer

Also known as

  • VAE behavioural modelling
  • Latent behaviour space
  • Generative behavioural archetypes

Categories

  • Analytics & Behaviour
  • Computational & Machine Learning