Overview
Unsupervised Machine Learning covers three topics. Clustering compares k-means, hierarchical and density-based methods and shows how to evaluate clusters when no labels exist. Principal component analysis (PCA) introduces the algorithm, its practical uses and how to interpret components and explained variance. Generative models cover the main types, their applications and how to evaluate them.
The module complements Module 4: supervised learning predicts a known target, while these methods discover structure, compress data or learn the data distribution itself. Results often feed back into supervised work as segment labels, reduced features or synthetic training data. The ideas also prepare you for Module 8, where generative adversarial networks are covered as a deep learning technique, and for Module 6, where cluster and component results must be communicated to people who never see the algorithm.
How it's tested
Expect method-choice scenarios where cluster shape, noise, dataset size or the need for a hierarchy decides the answer, plus interpretation items on silhouette scores, dendrograms, explained variance and component loadings. Generative-model questions focus on what these models learn and how they are evaluated.
-
Choose the clustering method
The scenario describes cluster shapes, outliers, whether k is known and data volume, and you choose k-means, hierarchical clustering or DBSCAN.
-
Interpret the output
A silhouette score, elbow plot, dendrogram or explained-variance table is described, and you draw the right conclusion about k or the number of components.
-
Generative or discriminative
A use case such as synthetic data, anomaly detection or imputation is described, and you identify the model type that learns the data distribution and how to evaluate it.
A team runs PCA on 40 unscaled variables, and the first component explains 92% of the variance with almost all its loading on annual revenue in dollars. The likely cause is scale: revenue's huge variance dominates because PCA maximizes variance. Standardizing the variables first lets components reflect correlation structure. Concluding that revenue is the only important variable would be the wrong reading.
Topics 3
Official topics from the Enterprise Big Data Framework Alliance outline, in outline order.
5.1 ClusteringK-means, hierarchical and density-based clustering, and evaluating clustering results.
- What it means
- Clustering groups observations so that members of a group are more similar to each other than to other groups. K-means assigns points to the nearest of k centroids; hierarchical clustering builds a tree of merges or splits shown as a dendrogram; density-based methods such as DBSCAN grow clusters from dense regions and label sparse points as noise.
- What to understand
- K-means needs k in advance, assumes compact, roughly spherical clusters of similar size, is sensitive to initialization (k-means++ helps) and to feature scale. Agglomerative clustering needs no k up front but its linkage choice (single, complete, average, Ward) shapes the result, and it scales poorly to very large n. DBSCAN needs eps and minPts, finds arbitrary shapes and noise, but struggles with clusters of very different density.
- How it shows up
- Points form irregular, elongated groups with scattered outliers and an unknown number of clusters, and you choose DBSCAN; or an elbow plot and silhouette scores are given and you choose k.
- Common confusion
- Treating a lower within-cluster sum of squares as always better; it keeps falling as k grows, so it cannot pick k alone.
5.2 Principal Component Analysis (PCA)Introducing PCA, the algorithm, practical applications and interpreting results.
- What it means
- PCA finds new orthogonal axes, the principal components, that are linear combinations of the original variables ordered by how much variance they capture. It is used for dimensionality reduction, visualization, noise reduction and removing multicollinearity before modeling.
- What to understand
- The algorithm centers (and usually standardizes) the data, computes the covariance matrix, and takes its eigenvectors as component directions and eigenvalues as variance captured. Choose the number of components from cumulative explained variance or a scree plot. Loadings show how each original variable contributes. PCA is linear and unsupervised, so the top components need not be the most predictive of a target.
- How it shows up
- An explained-variance table shows the first five components reach 91% and you choose five, or unscaled data produces a component dominated by one large-unit variable.
- Common confusion
- Treating PCA as feature selection: it creates new combined variables rather than keeping a subset of the originals.
5.3 Generative ModelsTypes of generative models, their applications and how to evaluate them.
- What it means
- A generative model learns how the data is distributed, P(X) or the joint P(X, Y), so it can generate new samples or score how likely an observation is. Discriminative models instead learn P(Y | X) directly. Types include Gaussian mixture models, naive Bayes, hidden Markov models, variational autoencoders, generative adversarial networks and autoregressive and diffusion models.
- What to understand
- Gaussian mixture models fit with expectation-maximization and give soft cluster memberships, unlike k-means' hard assignments. Applications include synthetic data, data augmentation, imputation and anomaly detection by flagging low-likelihood points. Evaluation uses held-out log-likelihood, AIC or BIC for model choice, Frechet Inception Distance for images, and for synthetic tabular data, fidelity, utility (train on synthetic, test on real) and privacy checks.
- How it shows up
- A hospital wants shareable data without exposing patients and you choose a generative model plus utility and privacy evaluation, or you must say why a GMM suits overlapping segments better than k-means.
- Common confusion
- Judging a generative model by how realistic a few samples look; it can memorize training records or miss whole regions of the data.
High-yield concepts
The terms, rules and formulas this domain turns into questions most often.
-
K-means objective and k-means++
K-means minimizes the within-cluster sum of squared distances to centroids by alternating assignment and centroid update steps.
- Remember
- Results depend on initialization; k-means++ spreads initial centroids and multiple runs reduce poor local optima.
- Tested as
- Method-choice and troubleshooting items.
- Trap
- Expecting identical clusters on every run without fixing the seed.
-
Elbow method and silhouette score
The elbow method looks for the k where adding clusters stops sharply reducing within-cluster sum of squares; the silhouette score, from -1 to 1, compares each point's cohesion with its separation from the nearest other cluster.
- Remember
- Silhouette near 1 means well separated, near 0 overlapping, negative suggests misassignment.
- Tested as
- Choosing k from given plots or scores.
- Trap
- Choosing the k with the lowest within-cluster sum of squares.
-
Linkage methods
Single linkage uses the closest pair between clusters, complete the farthest, average the mean distance, and Ward the increase in within-cluster variance.
- Remember
- Single linkage can chain into long straggly clusters; Ward favors compact clusters of similar size.
- Tested as
- Hierarchical clustering scenarios and dendrogram interpretation.
- Trap
- Assuming linkage choice does not change the clusters.
-
Dendrogram
A tree showing the order and distance at which clusters merge.
- Remember
- Cutting the tree at a height sets the number of clusters; long vertical gaps suggest natural cut points.
- Tested as
- Interpretation items ask how many clusters a cut produces.
- Trap
- Reading horizontal closeness of leaves as similarity instead of merge height.
-
DBSCAN
Density-based clustering with two parameters: eps (neighborhood radius) and minPts (points needed for a dense region); points are core, border or noise.
- Remember
- It finds arbitrary shapes and labels outliers as noise without needing k.
- Tested as
- Irregular shapes plus outliers point to DBSCAN.
- Trap
- Choosing it when clusters have very different densities, where one eps cannot fit all.
-
Internal vs. external cluster validation
Internal measures (silhouette, Davies-Bouldin, within-cluster sum of squares) use only the data; external measures (adjusted Rand index) compare with known labels.
- Remember
- Davies-Bouldin is better when lower; silhouette is better when higher.
- Tested as
- Evaluation questions with or without ground truth.
- Trap
- Reading a lower Davies-Bouldin value as worse.
-
Eigenvalues and explained variance
Each component's eigenvalue is the variance it captures; dividing by the total gives the explained-variance ratio.
- Remember
- Choose components by cumulative explained variance or the scree plot's elbow.
- Tested as
- Tables of explained variance appear in interpretation items.
- Trap
- Keeping every component, which removes the dimensionality reduction.
-
Standardizing before PCA
PCA maximizes variance, so variables with large units dominate unless standardized.
- Remember
- Standardize when variables are on different scales (equivalent to PCA on the correlation matrix).
- Tested as
- Diagnosis items where one variable dominates the first component.
- Trap
- Concluding a large-unit variable is the most important.
-
Loadings
The weights of the original variables in each principal component.
- Remember
- Large absolute loadings show which variables drive a component; signs show direction.
- Tested as
- Naming or interpreting components in business terms.
- Trap
- Treating a component as if it were one original variable.
-
Generative vs. discriminative models
Generative models learn P(X) or P(X, Y) and can sample new data; discriminative models learn P(Y | X) for prediction.
- Remember
- Naive Bayes and Gaussian mixtures are generative; logistic regression and SVMs are discriminative.
- Tested as
- Classification of model types and use-case matching.
- Trap
- Labeling logistic regression generative because it outputs probabilities.
-
Gaussian mixture model
Models data as a weighted mix of Gaussian components fit by expectation-maximization, giving each point a probability of belonging to each component.
- Remember
- Soft assignments and elliptical clusters, compared with k-means' hard, spherical ones; choose components with BIC.
- Tested as
- Overlapping-segment scenarios.
- Trap
- Treating it as identical to k-means.
-
Evaluating synthetic data
Synthetic data is judged on fidelity (matches real distributions), utility (models trained on it work on real data) and privacy (no near-copies of real records).
- Remember
- Train on synthetic, test on real is a standard utility check.
- Tested as
- Data-sharing and augmentation scenarios.
- Trap
- Assuming synthetic data is automatically private.
Side by side
Concepts candidates mix up, compared on the point that decides the answer.
K-means vs. hierarchical vs. DBSCAN
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| K-means | Assigns points to the nearest of k centroids | Needs k; best for compact, similar-size clusters; scales to large data | Using it on elongated shapes or with heavy outliers |
| Hierarchical (agglomerative) | Merges clusters step by step into a dendrogram | No k up front; shows nested structure; memory grows with n squared | Running it on millions of rows |
| DBSCAN | Grows clusters from dense neighborhoods | Arbitrary shapes; explicit noise; no k | Applying one eps to clusters of very different density |
PCA vs. feature selection
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| PCA | Builds new orthogonal components from all original variables | Unsupervised; maximizes variance; components are harder to interpret | Expecting it to keep the most predictive original features |
| Feature selection | Keeps a subset of the original variables | Often uses the target; preserves interpretability | Expecting it to remove multicollinearity as PCA does |
K-means vs. Gaussian mixture model
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| K-means | Hard assignment to nearest centroid | Spherical clusters; no membership probabilities | Forcing customers who fit two segments into one |
| Gaussian mixture model | Soft, probabilistic membership across Gaussian components | Elliptical clusters with their own covariance; generative | Forgetting it can also overfit with too many components |
Exam traps
Each trap is an answer that looks right. Learn the clue that should change your mind.
-
K-means by default
- What candidates pick
- K-means for GPS points that follow winding roads with scattered outliers.
- Why it looks right
- It is the best-known and fastest clustering method.
- Why it's wrong
- K-means assumes compact, roughly spherical clusters and forces every outlier into a cluster; DBSCAN follows arbitrary shapes and labels noise.
Clue to watch for Irregular or elongated shapes, outliers and an unknown number of clusters.
-
Lowest inertia picks k
- What candidates pick
- The k with the smallest within-cluster sum of squares.
- Why it looks right
- Lower error sounds better.
- Why it's wrong
- The sum keeps falling as k increases and reaches zero when every point is its own cluster; use the elbow or silhouette instead.
Clue to watch for A table lists within-cluster sum of squares for several k values.
-
Scale dominates PCA
- What candidates pick
- Conclude revenue is the key driver because it dominates the first component.
- Why it looks right
- The component loading is large and clear.
- Why it's wrong
- Without standardization, the variable with the largest variance in its own units dominates; the result reflects units, not importance.
Clue to watch for Variables are in very different units and one dominates the first component.
-
PCA for prediction by default
- What candidates pick
- Keep the top two components because they will be the most predictive features.
- Why it looks right
- They capture the most variance.
- Why it's wrong
- PCA ignores the target, so high-variance directions may carry little predictive signal; validate the downstream model.
Clue to watch for The goal is predicting a specific target, not visualizing or compressing.
-
Synthetic means private
- What candidates pick
- Share generated patient records freely because they are synthetic.
- Why it looks right
- No real record is shared directly.
- Why it's wrong
- Generative models can memorize and reproduce near-copies of training records; privacy must be tested, not assumed.
Clue to watch for Sensitive source data and an external sharing goal.
-
Hierarchical clustering at any scale
- What candidates pick
- Agglomerative clustering on 50 million transactions.
- Why it looks right
- It does not need k and gives a dendrogram.
- Why it's wrong
- Standard agglomerative methods need pairwise distances, so memory and time grow at least with n squared; use a scalable method or cluster a sample.
Clue to watch for Very large row counts are stated alongside a hierarchical option.
Answer framework
Work through these steps for any unsupervised-learning scenario on the exam.
- Define the goal clearly: discover segments, compress variables, visualize, detect anomalies or generate new data that resembles the source.
- Check the data's scale and units, and standardize before any distance-based clustering or PCA when variables differ in scale.
- Read the shape clues, including cluster form, outliers, known or unknown k, overlap between groups and row count, to pick k-means, hierarchical, DBSCAN or a mixture model.
- Choose k or the number of components with the elbow, silhouette, dendrogram cut or cumulative explained variance, not a single raw error figure.
- Validate the result against its purpose, using downstream accuracy, interpretability, or fidelity, utility and privacy for generated data.
Practice question
Want more like this? The EBDS practice test has questions for every domain.
Checklist
0 of 13
Tick each one you can do without notes. Your ticks stay in this browser.
Study strategy
Start with topic 5.1 by running k-means, agglomerative clustering and DBSCAN on three toy datasets: round blobs, concentric moons and blobs with outliers. Seeing each method fail on the wrong shape is the fastest way to answer method-choice questions. Then practice choosing k from elbow plots and silhouette scores, and cutting a dendrogram.
Move to topic 5.2 and run PCA on a mixed-unit dataset twice, with and without standardization. Compare the explained-variance tables and loadings; this single exercise covers the most common PCA traps. Practice naming the first two components from their loadings.
Topic 5.3 is commonly underestimated because generative models sound like a deep learning topic. Focus on the generative versus discriminative distinction, Gaussian mixtures versus k-means, and evaluation of synthetic data. Afterwards, take timed scenario sets that mix clustering choice and PCA interpretation, then continue to Module 6, where you learn to present segments and components to non-technical audiences.
FAQ
Is k-means the main clustering method on the EBDS exam?
The syllabus names k-means, hierarchical and density-based clustering plus evaluation, so expect questions that require choosing among them rather than only applying k-means.
How many questions come from unsupervised learning?
No module weights are published. An even split of 80 questions suggests about nine; treat that as a planning estimate.
How do I choose the number of clusters?
Use the elbow in within-cluster sum of squares, silhouette scores, a dendrogram cut or BIC for mixture models, and check that the clusters are meaningful for the business purpose.
Should I standardize before PCA?
Yes when variables are on different scales; otherwise the largest-unit variables dominate the components regardless of their importance.
What is the difference between PCA and feature selection?
PCA creates new components from combinations of all variables, while feature selection keeps a subset of the original variables and preserves their meaning.
Why are generative models in the unsupervised module?
Most learn the distribution of the data without needing labels, which lets them generate samples, score likelihood for anomaly detection or impute missing values.
How does this module connect to deep learning?
Variational autoencoders and generative adversarial networks are deep generative models; Module 8 covers how GANs are built and trained.
Other Data & Analytics certifications
Where the same subject shows up in other exams. Cards marked "Same subject" open the matching domain.
Test your Domain 5 knowledge
You have covered clustering choice and validation, PCA and its interpretation, and generative models. Check whether you can pick the right unsupervised method from the clues in a scenario.