Two-stage clickstream persona construction

Description

Cluster session clickstreams into common workflows, then fit a mixture model over each user's workflow counts to derive behavioural personas from product telemetry.

Telemetry events are split into per-session clickstreams, dropping those under 20 clicks. Pairwise Jaccard similarity of unique clicks feeds Ward hierarchical clustering, giving K common workflows that are labelled with experts from top clicks, length, duration and key steps. Each user's counts of each workflow are then modelled as a multinomial given a latent persona; a mixture of M multinomials is fitted by EM, and users are assigned to their most likely persona. Personas are read from workflow probability vectors and a user-by-workflow heat map, and experts compare them with legacy personas. At Platfora, 3.5 million clicks from 2,400 users yielded ten workflows and five personas. Limitations: Jaccard on unique clicks ignores order and repetition, K and M were chosen by judgement, and the personas show behaviour, not goals, so they need triangulation with interviews.

Attributes

Loss type
Latent
Effort
Weeks

Steps

  1. Collect opt-in UI event telemetry with anonymised user identifiers.
  2. Split events into per-session clickstreams and drop those under 20 clicks.
  3. Compute a Jaccard distance matrix between clickstreams and run Ward hierarchical clustering; choose K workflows with the dendrogram and expert review.
  4. Label workflows from top clicks, length, duration and key-step frequencies, dropping stray clusters such as internal testing.
  5. Count each user's workflows, keeping users with at least five clickstreams.
  6. Fit a mixture of M multinomials by EM and assign users to their most likely persona, checking fit statistics where possible.
  7. Describe personas by workflow probabilities and have UX, design and product experts compare them with existing personas.

Sources

  • Data-driven Personas (Zhang et al., 2016)

Also known as

  • Two-stage clickstream persona construction (workflow clustering then multinomial mixture of users)
  • Clickstream personas
  • Telemetry-based behavioural personas
  • Workflow-mixture personas

Categories

  • Analytics & Behaviour
  • Computational & Machine Learning