BSF Stage 1 — Wavelet Decomposition

Description

Decomposes user behaviour time series into components at multiple temporal scales, separating session-level activity from weekly routines and lifecycle trends.

Wavelet decomposition is the first stage of the Behavioral Spectral Framework playbook. User activity is represented as a time series, for example events per hour or per day for each user, and decomposed into components at several temporal resolutions, analogous to a discrete wavelet transform. Unlike a single Fourier transform, wavelets localise patterns in time, so a weekly routine that begins in month two or a burst of activity around a billing event can be detected where it occurs. The stage turns undifferentiated activity counts into separate layers that later stages map, filter and model. Its rigour depends on consistent time binning, sufficient series length relative to the slowest scale studied, handling of missing periods, and checking that the components sum back to the original signal.

Attributes

Loss type
Community-proposed
Stage
Post-launch
Effort
Days

Steps

  1. Choose the unit of analysis (user, account or cohort) and a consistent time bin such as hourly or daily event counts.
  2. Ensure each series spans at least three to four cycles of the slowest rhythm to be studied.
  3. Handle gaps explicitly, distinguishing true inactivity from missing data.
  4. Apply a multi-level wavelet decomposition (for example a discrete wavelet transform with a Haar or Daubechies basis).
  5. Verify that reconstructed components sum to the original series within tolerance.
  6. Store components per scale for each unit for use in frequency mapping.

Sources

  • wavelet.pdf

Also known as

  • Behavioural wavelet decomposition

Categories

  • Community-Proposed Frameworks