Hidden Markov Models for Navigation Analysis

Description

Infers latent user states (exploring, task-focused, confused, recovering) from clickstream sequences and estimates transitions between them.

An HMM models observed sequences as emissions from hidden states with transition probabilities. Given clicks, page visits or actions, inference recovers the most likely state sequence (Viterbi) and transition matrix. The treatise poses a classic question: is a long session engagement or confusion? An HMM trained on labelled sessions can separate these as different hidden states with different observable signatures, then label new sessions. This turns ambiguous metrics such as session length into state-based measures (time in 'confused' state). Model selection uses information criteria and interpretability, and states should be validated against qualitative evidence such as recordings or think-aloud.

Attributes

Loss type
Unmigrated
Stage
Post-launch
Effort
Weeks

Steps

  1. Define the observation alphabet (event types, page categories, dwell bins).
  2. Optionally label a sample of sessions with states via observation or recordings.
  3. Fit HMMs with different numbers of states using Baum-Welch or supervised estimation.
  4. Select the model by BIC, held-out likelihood and interpretability.
  5. Decode sessions with Viterbi and compute time-in-state metrics.
  6. Validate states qualitatively and track state metrics across releases.

Sources

  • Lost Methods in UX — A Complete Treatise (Gopinath)
  • Rabiner LR (1989) A tutorial on hidden Markov models and selected applications in speech recognition. Proc IEEE 77(2):257-286
  • Baum LE & Petrie T (1966) Statistical inference for probabilistic functions of finite state Markov chains. Ann Math Stat 37

Also known as

  • HMM
  • Latent state sequence modelling

Categories

  • Analytics & Behaviour
  • Computational & Machine Learning