A glossary of tiny lectures, part 10

Revision: one machine, many names

By Tim Menzies, published 2026-11-09, updated 2026-11-09

Summary: No new terms here. This page is the map: what each week bought, how the pieces stack into one machine, and the claim the whole course exists to test — that most of the "different" tools in a data mining textbook are the same tool, wearing hats. Read this before the final, and before you write up either project.

The stack, in one page

Every week added one layer, and every layer used only the ones below it.

week	what it added	what that bought
1	columns that summarize as they read (Num, Sym, add, mid, div)	O(1) memory, and summaries that subtract
2	tables, distance, disty (gap to heaven)	one number for "how good is this row"
3	fastmap poles, projection, recursive halving	structure with no labels, and an index
4	cuts scored by expected diversity; trees; XAI	a model small enough to argue with
5	budgets, acquire, restart, holdout	good rows for about 45 labels, honestly scored
6	cohen, cliffs, ks, ranks, domination, wins	the right to say "this one is better"
7	knn, anomaly, naive Bayes, kmeans, kpp	the textbook chapters, a dozen lines each
8	ga, de, sa, local search, race	fifty years of search, and its price tag
9	the model seam, DTLZ	goals that cost money, and known right answers

The one machine. Read down this picture and every week is in it:

rows (x cheap, y expensive)
  |
  +-- distx -----> poles -----> projection -----> halve
  |                                                  |
  |                                            recurse = tree
  |                                                  |
  +-- disty (buys labels) --> better half only --> acquire
                                                     |
                              leaf -> mean = regression
                              leaf -> mode = classification
                              far from leaf = anomaly / alert
                              read the tree  = summarization / XAI
                              mutate + re-drop = forecast / what-if

The synonyms claim

Buse & Zimmermann's analytics grid has nine boxes; the explanation literature has its own list of triggers. The claim is that those lists collapse:

they call it	we call it
classification	the leaf's mode
regression, forecasting	the leaf's mean
optimization, tuning	keep the better half, recurse
alerts, anomaly detection	far outside the leaf's own spread
benchmarking	anomaly detection, with a reference population
simulation, what-if, goals	mutate x, re-drop, read the new leaf
summarization, overlays	read the tree
privacy	publish the corner (prototypes of the columns that matter)
explanation	the tree is the explanation

How you can prove this wrong. The claim is falsifiable, and the undergraduate project is the experiment: build the skills in order, tag each step, and plot the percentage of NEW code per step. If reuse is real the bars fall. If they stay flat, this course's architecture lost — and an honest flat histogram scores full marks, because failures are findings.

The rival is easy to run: ask an LLM for each skill cold, one fresh prompt per skill, with no instruction to reuse anything. The prediction to beat is a thousand lines of new glue per skill, ten separate programs, each hooked into a large library.

Ideas worth carrying out of here

x is cheap, y is expensive. Almost everything in this course follows from that one asymmetry.
Weak indicators, hedged. A cheap heuristic is evidence, not truth: ask it more times, or act on it more gently.
Before optimizing a computation, ask how little data it needs. Poles from 128 rows; cuts from a sample; prototypes instead of rows.
Support, not just purity. A split that isolates one row explains nothing.
Say what it cost. A score without its label count, or without saying pure-vs-generalize, is not a result.
Ties are the normal case. Rank 0 is usually a set. So is a Pareto front.
Know when you do not know. The certification envelope is three lines and it is the difference between a useful model and Crater.
Explanations must be actionable. Not "tires are important" — bald tires.

Where to look things up. Principles and weeks 0-2: gloss1. The Lua machinery under the demos: gloss2. Clustering: gloss3. Cuts and trees: gloss4. Active learning: gloss5. Statistics: gloss6. Applications: gloss7. Optimizers: gloss8. Models and the seam: gloss9. The full text of every entry lives in one file: glossary.md.

Buse & Zimmermann: Information needs for software development analytics. Raymond P.L. Buse, Thomas Zimmermann. ICSE 2012.

Hoffman, Mueller, Klein & Litman: Metrics for explainable AI: challenges and prospects. Robert R. Hoffman, Shane T. Mueller, Gary Klein, Jordan Litman. arXiv:1812.04608, 2018.