ch03 — Gen2 Classical Machine Learning (2000–2016)

Companion notebook: notebooks/ch03_gen2_classical_ml.ipynb — UCR discords with stumpy

Density and boundary methods: LOF, OC-SVM, KNN

Embed the series as subsequence window vectors (embed_windows) and any generic outlier detector applies:

• KNN / Sub-KNN: distance to the k-th nearest train window. Simple, and very strong on univariate data.
• LOF: density relative to the local neighborhood — beats KNN when density varies across the space.
• OC-SVM: learns a minimal boundary around normal data; \(\nu\) controls the fraction allowed outside.

Isolation Forest — "what isolates easily is anomalous"

In trees built from random axes and random splits, anomalies isolate at shallow depth. No density estimation, O(n log n), robust in high dimensions — the most common practical default. It is stochastic, so this repository evaluates it with 3 seeds (stochastic: true).

Matrix Profile — discovering discords

The profile of minimum z-normalized Euclidean distances between subsequences, \(MP_i = \min_{j} d(w_i, w_j)\). A discord (a subsequence whose nearest neighbor is still far) matches the definition of a collective anomaly exactly.

• One parameter (window length) — close to parameter-free
• Needs no train split (test self-join) — genuinely unsupervised
• Repeated anomalies can hide each other (the twin freak problem)

The UCR Anomaly Archive is essentially a discord-discovery benchmark, which is why MP is so strong there. We use stumpy (BSD-3) as a pip dependency (matrix_profile).

Where is MERLIN?

MERLIN (discord discovery at all lengths) is deferred: the public reference implementation has no clear license (CLAUDE.md §10-2 — no license, no copying). Recorded in THIRD_PARTY_NOTICES.md.