Module 3: Clustering & Anomalies

Comprehensive vocabulary flashcards covering unsupervised learning, k-means clustering mechanics, data transformation, model diagnostics, dimensionality reduction via PCA, and anomaly detection.

Supervised Learning

An analytical approach where a model learns from historical training data containing a known answer column (target labels) and is evaluated using an accuracy score.

Unsupervised Learning

An analytical approach operating on data with no answer column or target labels, designed to discover inherent structure by determining which rows resemble other rows.

$k$-means Clustering

An algorithm that groups data points into $k$ clusters by dropping $k$ markers at random, assigning each observation to the nearest marker, moving each marker to the centroid of its assigned points, and repeating until convergence.

$$k$$

The user-specified input parameter defining the number of clusters to form in $k$-means; it is supplied by the analyst rather than determined automatically by the algorithm.

Standardization

The preparation technique of converting each feature column into standard deviations from its own mean so that attributes measured on different scales have equal influence in distance calculations.

Log Transformation

Converting skewed numerical data (such as monetary spend) to a logarithmic scale so the column measures order of magnitude, preventing single massive outliers from claiming an entire cluster marker.

Elbow Method

A heuristic tool for selecting $k$ that plots cluster tightness against the number of clusters to identify the bend where incremental improvements in tightness slow down.

Silhouette Score

A geometric metric ranging from $-1$ to $1$ that measures how much closer an observation is to its own cluster compared to the next nearest cluster, where values above $0.5$ indicate reasonable structure.

Coverage

The proportion of genuinely problematic or irregular entities in a population that a selected partition successfully places before an audit or review team.

Wasted Effort

The proportion of reviewed entities that turn out to be completely fine, expending audit capacity and poking supplier relationships without finding issues.

Principal Component Analysis (PCA)

A mathematical dimensionality reduction method that finds orthogonal directions of maximum variance to flatten high-dimensional data onto a two-dimensional map.

Centroid

The geometric center point of a cluster in $k$-means, calculated as the mathematical average across all feature dimensions of every observation assigned to that cluster.

Anomaly

An observation that sits unusually far from its assigned cluster centroid because its specific combination of attributes across multiple dimensions does not fit any standard profile.

Random Restarts

The procedure of running the $k$-means algorithm multiple times from different initial random marker positions and selecting the tightest overall result to avoid settling into suboptimal local arrangements.