Exercise 2: Why the Employee-Clustering Finding Is Genuinely New — Possible Solution
====================================================================

WHAT ml1-5's OWN SUPERVISED MODEL WAS ACTUALLY BUILT TO DO
------------------------------
Per ml1-5, the logistic regression model was fit specifically to
predict left_company — a single, pre-specified target variable — from
department, age, years_at_company, and salary. Its entire job, by
design, was answering one particular pre-formulated question: given
these features, how likely is this specific outcome (leaving)?

WHY IT COULD NEVER HAVE SURFACED THE UNDERPAID-VETERAN PATTERN ON ITS OWN
------------------------------
The underpaid-long-tenure pattern this chapter's own clustering found
has nothing intrinsically to do with left_company at all — it's a
pattern purely among age, years_at_company, and salary themselves, with
no reference to whether those particular employees left or stayed.
ml1-5's own model was never asked, and had no mechanism, to look for
groupings among the FEATURES independent of the target it was trained
to predict — its entire fitting process (per ml1-5's own coefficient-
learning description) is organized entirely around predicting
left_company, not around describing structure that exists among the
inputs regardless of that outcome.

WHAT THIS CHAPTER'S OWN CLUSTERING DID DIFFERENTLY
------------------------------
Per this chapter, clustering was run using only age, years_at_company,
and salary — with left_company never included as an input at all — and
with no target of any kind to predict. Because the algorithm's job was
simply "find groups of similar employees," rather than "predict this
one specific outcome," it was free to discover ANY structure present
among those three features, including a structure — long tenure paired
with comparatively low pay — that has no necessary connection to
attrition specifically.

WHY THIS IS EXACTLY WHAT "UNSUPERVISED LEARNING IS GENUINELY FOR"
------------------------------
Per this chapter, "ml1-5's own logistic regression could only ever
answer the specific question it was given... This chapter's clustering
wasn't given a question at all — it surfaced a group nobody had
explicitly asked about." A supervised model is fundamentally
constrained to whatever single question its target variable encodes;
an unsupervised model has no such constraint, and can surface real
structure that nobody thought to formulate as an explicit question in
the first place — precisely the underpaid-veteran cluster, which was
never something anyone set up left_company (or any other target) to
detect.

WHY THIS WORKS AS AN ANSWER
------------------------------
It explains precisely what ml1-5's own model was scoped to predict (one
specific target, left_company) and why that scoping structurally
prevented it from ever surfacing a pattern unrelated to that target, and
contrasts this with clustering's own lack of any target at all, which
is exactly what allowed it to discover a genuinely new, previously
unasked-about pattern.