Exercise 1: Why K-Means's Fitting Process Is Genuinely Different — Possible Solution ==================================================================== WHAT ml1-3's LEAST-SQUARES FITTING ACTUALLY DOES ------------------------------ Per ml1-3, least-squares fitting minimizes the total squared distance between each PREDICTED value and each ACTUAL, known value — the model is being continuously corrected against a real, external answer (the true price) that exists independently of the model and is provided during training. WHAT ml1-7's ENTROPY-DRIVEN SPLITTING ACTUALLY DOES ------------------------------ Per ml1-7, a decision tree picks each split by measuring information gain against KNOWN LABELS — whether an employee actually left or stayed, a real, external, provided answer the tree checks its own candidate splits against at every single node. WHAT K-MEANS DOES INSTEAD, PER THIS CHAPTER'S OWN FIVE STEPS ------------------------------ Per this chapter's own step list: "randomly place k initial centroids... assign every data point to its nearest centroid... recompute each centroid as the average position of every point now assigned to it... repeat... until the centroids stop moving." At no point in this process does the algorithm compare anything against a real, external, known answer — there is no true cluster label anywhere in the data for it to check its own guesses against. Instead, it starts from an essentially arbitrary random guess (the initial centroid positions) and repeatedly refines that guess against ITSELF — each round's new centroid positions are computed from the previous round's own cluster assignments, which were themselves based on the round before that. WHY THIS IS THE SPECIFIC SENSE IN WHICH IT'S "GENUINELY DIFFERENT" ------------------------------ Per this chapter, "no least-squares fit against a known target..., no entropy-driven split search against known labels... k-means iteratively refines its own evolving guess about where the clusters are, with nothing external to check itself against." Both of the earlier algorithms have an external ground truth guiding every step of fitting; k-means has no such ground truth available at all — its own "correctness" is defined entirely relative to its own current guess converging to a stable configuration, not relative to any independently known right answer. WHY THIS DIRECTLY REFLECTS THIS CHAPTER'S OWN UNSUPERVISED FRAMING ------------------------------ This mechanical difference is the concrete, algorithmic version of this chapter's own opening claim: unsupervised learning has "no known answer to train against." Least squares and entropy-driven splitting are both fundamentally SUPERVISED processes at the level of their own fitting mechanics — always checking against a real external target. K-means's own iterative self-refinement process has no equivalent external target anywhere in its design, which is precisely why it belongs to a genuinely different category of algorithm, not simply a variation on the earlier ones. WHY THIS WORKS AS AN ANSWER ------------------------------ It contrasts precisely what ml1-3 and ml1-7 check their own fitting process against (a real, known, external answer) with what k-means checks its own fitting process against (nothing external at all, only its own prior guess), explaining exactly why this chapter calls the mechanism "genuinely different" rather than merely a new example of the same underlying idea.