K-Means, DBSCAN and Hierarchical Clustering: Which Fits Evidence Organisation?

Understand what each algorithm asks you to choose

K-Means asks for a number of clusters and assigns points around centroids. DBSCAN uses a density neighbourhood and can leave points as noise. Hierarchical methods build groups through a linkage rule. These differences affect how outliers, uneven groups and ambiguous evidence are treated.

The twelve broad issue families are not a reason to set k to twelve. Families describe subjects; concrete issue counts vary by case. Likewise, DBSCAN noise does not mean irrelevant evidence. A single important request may have no nearby neighbours and still deserve prominent review.

Compare assumptions before fitting real data

This configuration sketch makes the experiment’s differences explicit. The values are starting points for a synthetic exercise, not MailTrace settings. Real comparisons need compatible normalisation, a stated distance metric and validation data for parameter selection.

{
  "kmeans": {"clusters": "selected on validation data", "unassigned": false},
  "dbscan": {"radius": "selected on validation data", "unassigned": true},
  "complete_link": {"merge_rule": "all cross-pairs compatible", "unassigned": "singletons"}
}

Use evidence-level outcomes to select a method

Compare false merges, missed connections, singleton rate and reviewer correction effort. Include cases with repeated templates and multiple requests in the same thread. Generic compactness scores can be diagnostic, but they do not establish that the grouping preserves the distinct matters a user needs to pursue.

Run each candidate offline against fixed source representations before considering a product change. Save the algorithm version and inspect disagreements with the current method. Any new candidate generator should feed the existing review workflow, preserve source membership and leave the user able to reject or reverse its suggestions.