Cluster from the notes; do not pour them into a ready taxonomy
Aliases: bottom-up affinity diagram · KJ method induction · a priori category dump
What it is
A wall already labeled Experience / Feature / Process, with notes merely slotted in, is not an affinity diagram. An affinity diagram (from Kawakita’s KJ method) starts with single observations and lets clusters form by how the contents sit next to each other; names come after the clusters exist. A ready taxonomy pours material into existing bins. Bottom-up induction lets the bins grow out of the material. Both produce a wall of notes. Only one of them is evidence of structure in the data.
Why it happens
People protect a structure once it has been spoken. Columns on the board force every subsequent note to answer “which column?” rather than “which other notes say the same thing?” Observations that will not fit get forced, rewritten until they fit, or quietly dropped. Working bottom-up cuts that shortcut: similar quotes and events are placed together first, and the similarity rule is articulated only in that placement. Org charts, product modules, and pre-written research questions are strong prior grids; they translate “how people lived it” back into “how we are staffed.”
Studying it
Take one set of observation notes and compare “sort into a supplied frame” with “no names yet, only proximity moves.” Count leftovers, items rewritten in order to land, and overlap between final clusters and the initial frame. Near-total overlap is a sign that induction never happened. Swap materials across two teams: one receives department columns, the other shuffled source quotes; see which observations survive only in the latter. Outcomes: leftover rate, forced-assignment count, and whether a non-participant can retell a cluster in the original wording.
Where it stops holding
A tested codebook, a compliance checklist, or a longitudinal scheme you must reuse is a legitimate top-down sort—just do not call it affinity work. When time is extreme and the material was already produced from one interview guide, bottom-up clustering may simply mirror the guide. In participatory workshops, stakeholders arrive with organizational language; unless “stick it by department” is explicitly closed, the wall becomes a handsome org chart rather than an observational structure.
Applying it
- One note, one identifiable event or quote; ban column-words such as “UX issue” at the writing stage.
- Start with no columns. Move neighbors in silence until small clusters stabilize, then argue about names.
- If a note can join a cluster only by product-module or department name, pull it back to the unsorted pile.
- At close, audit: if every cluster maps one-to-one onto the pre-session taxonomy, redo a round and keep the unplaceable notes as counterexamples.