Question 1

What is the main difference between supervised and unsupervised learning?

Accepted Answer

Unsupervised learning works with unlabeled data, seeking to discover hidden structures or patterns without a predefined target variable. Unlike supervised learning which predicts a known value (label), unsupervised learning explores data to find natural groups, reduce dimensionality, or detect anomalies. Algorithms like K-Means, PCA, or DBSCAN are typical examples of unsupervised learning.

Question 2

How does the K-Means algorithm work to partition data?

Accepted Answer

K-Means is an iterative algorithm that partitions data into K clusters. It randomly initializes K centroids, then alternates between two steps: assigning each point to the nearest centroid (assignment step) and recalculating centroid positions as the mean of assigned points (update step). The algorithm converges when assignments no longer change or after a maximum number of iterations.

Question 3

Which method should be used to determine the optimal number of clusters K in K-Means?

Accepted Answer

The elbow method plots inertia (sum of squared distances between each point and its centroid) against K. The point where the curve forms an elbow indicates optimal K, as beyond it adding clusters no longer significantly improves inertia. This method is complemented by silhouette score to validate cluster quality.

Question 4

What does the silhouette score measure in the context of clustering?

Accepted Answer

Silhouette score measures clustering quality by comparing intra-cluster cohesion (average distance to points in same cluster) and inter-cluster separation (average distance to nearest cluster). A score close to 1 indicates well-defined and separated clusters, close to 0 indicates overlapping clusters, and negative indicates poor point assignment.

Question 5

What is the range of silhouette score values and how to interpret a score of 0.7?

Accepted Answer

Silhouette score ranges from -1 to 1. A score of 0.7 is considered good and indicates well-defined clusters with good separation. General interpretations are: above 0.7 (excellent structure), between 0.5 and 0.7 (reasonable structure), between 0.25 and 0.5 (weak structure), below 0.25 (no clear structure or incorrect assignment).

Question 6

What major limitation of K-Means makes the algorithm unsuitable for non-spherical cluster shapes?

Accepted Answer

K-Means uses Euclidean distance to centroid as assignment criterion, which implicitly assumes clusters are spherical (isotropic) and of similar sizes. The algorithm cannot capture elongated shapes, crescents, or variable densities. For such cases, density-based algorithms like DBSCAN are more appropriate.

Question 7

How does agglomerative hierarchical clustering build a hierarchy of clusters?

Accepted Answer

Agglomerative hierarchical clustering uses a bottom-up approach: each point starts as its own cluster, then the closest clusters are iteratively merged according to a linkage method (single, complete, average, Ward). The process continues until one cluster remains. The result is visualized by a dendrogram that allows choosing the number of clusters by cutting at different heights.

Question 8

Which linkage method to use in hierarchical clustering to minimize intra-cluster variance?

Accepted Answer

Ward linkage minimizes the increase in intra-cluster variance at each merge. It produces compact clusters of similar sizes, similar to K-Means. Other methods are: single linkage (minimum distance, noise sensitive), complete linkage (maximum distance, compact clusters), and average linkage (average distance, compromise between both).

Question 9

How to use a dendrogram to determine the number of clusters in hierarchical clustering?

Accepted Answer

The dendrogram visualizes successive merges with height representing distance between merged clusters. To determine the number of clusters, cut the dendrogram horizontally at a chosen height. The general rule is to cut where vertical branches are longest (large distance before next merge), indicating a natural separation in the data.

Question 10

What are the two main hyperparameters of DBSCAN and their role?

Accepted Answer

DBSCAN uses two hyperparameters: epsilon (eps) which defines the neighborhood radius around each point, and min_samples which defines the minimum number of points required in that neighborhood for a point to be considered a core point. A point with at least min_samples neighbors in its epsilon-neighborhood forms the core of a cluster.

Question 11

What major advantage of DBSCAN distinguishes it from K-Means for outlier detection?

Accepted Answer

DBSCAN automatically identifies noise points (outliers) as points that are neither core points nor border points. These points cannot be reached from any core point within an epsilon neighborhood and are labeled as -1. K-Means, on the other hand, forcibly assigns every point to a cluster, even outliers, which can distort centroids.

Question 12

In which case is DBSCAN preferable to K-Means for clustering?

Accepted Answer

DBSCAN is preferable when clusters have arbitrary shapes (non-spherical), similar densities, and when the number of clusters is unknown. It also excels with noisy data or outliers. However, DBSCAN struggles with clusters of very different densities and is sensitive to the choice of epsilon and min_samples hyperparameters.

Question 13

What is the main objective of Principal Component Analysis (PCA)?

Accepted Answer

PCA is a dimensionality reduction technique that transforms data into a new coordinate system where axes (principal components) capture maximum variance. The first component captures most variance, the second captures most after orthogonal projection, and so on. This allows reducing the number of features while retaining most of the information.

Question 14

How to determine the number of principal components to retain with PCA?

Accepted Answer

The choice of component count is typically made by analyzing cumulative explained variance ratio. A common rule is to retain components that explain 95% of total variance. The scree plot (eigenvalues graph) also helps identify the elbow where adding components no longer provides significant variance gain.

Question 15

Why is it important to standardize data before applying PCA?

Accepted Answer

PCA maximizes variance, so features with large scales would dominate principal components. Standardization (mean 0, std 1) puts all features on equal footing. Without standardization, a feature in thousands (e.g., salary) would overshadow a feature in units (e.g., age), even if the latter is more informative for data structure.

Question 16

What is the main difference between PCA and t-SNE for data visualization?

Accepted Answer

PCA is a linear technique that preserves global structure and large distances, while t-SNE is non-linear and optimized to preserve local neighborhood relationships. t-SNE excels at revealing clusters and local structures in high-dimensional data, but doesn't preserve global distances. PCA is deterministic and reproducible, t-SNE is stochastic.

Question 17

What is the role of the perplexity parameter in t-SNE?

Accepted Answer

Perplexity in t-SNE controls the effective number of neighbors considered for each point, influencing the balance between local and global structure. Low perplexity (5-10) emphasizes small local structures, high perplexity (30-50) reveals more global patterns. Recommended values are typically between 5 and 50, with 30 as a common default.

Question 18

What advantages does UMAP offer over t-SNE for dimensionality reduction?

Accepted Answer

UMAP is significantly faster than t-SNE, especially on large datasets, and better preserves global data structure. UMAP can be used as preprocessing for machine learning because it creates a reproducible transformation applicable to new data (unlike t-SNE). It also supports custom distances and reduction to more than 2-3 dimensions.

Question 19

Which key UMAP parameter controls the balance between local and global structure preservation?

Accepted Answer

The n_neighbors parameter in UMAP controls local neighborhood size used to build the representation. A low value (5-15) preserves more local structure and reveals micro-clusters, while a high value (50-200) captures more global topology. The min_dist parameter controls cluster compactness in the projected space.

Question 20

Why should distances between clusters not be interpreted in a t-SNE visualization?

Accepted Answer

t-SNE optimizes local neighborhood preservation, not global distances. Two clusters far apart in original space may appear close in t-SNE, and vice versa. Cluster sizes can also be misleading as t-SNE adjusts scales locally. Only the existence of separate clusters is reliable, not their relative position or comparative size.

Question 21

In what order should PCA and K-Means be applied to improve clustering on high-dimensional data?

Accepted Answer

Applying PCA before K-Means is generally beneficial for high-dimensional data. PCA reduces noise and redundancy, mitigates the curse of dimensionality where distances become less meaningful, and speeds up K-Means. Principal components capture maximum variance directions that are often most relevant for cluster structure.

Question 22

How does K-Means++ initialization improve the standard K-Means algorithm?

Accepted Answer

K-Means++ selects initial centroids probabilistically, favoring points far from already chosen centroids. This smart initialization avoids the problem of initial centroids being too close, which can lead to convergence to a suboptimal local minimum. K-Means++ produces more stable results and generally converges faster than random initialization.

Unsupervised ML

What is the main difference between supervised and unsupervised learning?

Answer

How does the K-Means algorithm work to partition data?

Answer

Which method should be used to determine the optimal number of clusters K in K-Means?

Answer

What does the silhouette score measure in the context of clustering?

What is the range of silhouette score values and how to interpret a score of 0.7?

What major limitation of K-Means makes the algorithm unsuitable for non-spherical cluster shapes?

Master Data Science & ML for your next interview