Machine Learning Unsupervised Learning 5 — Questions and Answers
Question 1: Which problem can arise when using K-Means with very high-dimensional data?
- The curse of dimensionality makes Euclidean distances less meaningful (Correct answer)
- K-Means becomes a supervised algorithm
- Convergence is guaranteed in one iteration
- Cluster centroids cannot be computed
Correct answer: The curse of dimensionality makes Euclidean distances less meaningful
In high dimensions, distances between points become increasingly similar (concentration of measure), making it hard for K-Means to distinguish meaningful clusters.
Question 2: What is 'spectral clustering' and what advantage does it have?
- It uses eigenvectors of a similarity graph Laplacian to cluster data of non-convex shapes (Correct answer)
- It clusters data using the frequency spectrum of audio signals
- It applies PCA before K-Means for speed
- It uses hierarchical linkage with spectral distances
Correct answer: It uses eigenvectors of a similarity graph Laplacian to cluster data of non-convex shapes
Spectral clustering constructs a similarity graph, computes its Laplacian eigenvectors, then clusters the low-dimensional embedding, allowing non-convex cluster detection.
Question 3: In Self-Organizing Maps (SOMs), what is the 'neighborhood function' responsible for?
- Updating weights of neurons near the winning neuron to preserve topological structure (Correct answer)
- Selecting the number of output neurons
- Normalizing input features before training
- Computing the reconstruction error
Correct answer: Updating weights of neurons near the winning neuron to preserve topological structure
The neighborhood function ensures that neurons near the winning (best-matching) unit are also updated, causing the map to preserve topological relationships of the input data.
Question 4: What is 'co-clustering' (biclustering) in unsupervised learning?
- Simultaneously clustering both rows and columns of a data matrix (Correct answer)
- Clustering two separate datasets together
- Running K-Means twice with different initializations
- Merging two dendrogram branches simultaneously
Correct answer: Simultaneously clustering both rows and columns of a data matrix
Biclustering finds subsets of rows and columns in a matrix that exhibit coherent patterns, commonly used in gene expression analysis.
Question 5: Why is random initialization in K-Means a concern, and how does K-Means++ address it?
- Random init can lead to poor local minima; K-Means++ spreads initial centroids to improve convergence (Correct answer)
- Random init is too slow; K-Means++ uses greedy centroid selection for speed
- Random init ignores outliers; K-Means++ removes them first
- Random init fixes K; K-Means++ selects K automatically
Correct answer: Random init can lead to poor local minima; K-Means++ spreads initial centroids to improve convergence
K-Means++ selects initial centroids with probability proportional to their distance from already-chosen centroids, reducing the chance of converging to a bad local minimum.
Question 6: What does the Davies-Bouldin Index measure in cluster evaluation?
- The ratio of within-cluster scatter to between-cluster separation; lower is better (Correct answer)
- The number of clusters relative to data size
- The percentage of variance explained by clusters
- The probability that two random points share the same cluster
Correct answer: The ratio of within-cluster scatter to between-cluster separation; lower is better
The Davies-Bouldin Index computes the average similarity ratio of each cluster with its most similar cluster; a lower score indicates better-defined, well-separated clusters.
Question 7: Which unsupervised technique is most suitable for finding association rules like 'customers who buy X also buy Y'?
- Apriori algorithm (frequent itemset mining) (Correct answer)
- K-Means clustering
- PCA
- Autoencoder
Correct answer: Apriori algorithm (frequent itemset mining)
The Apriori algorithm mines frequent itemsets from transaction data and derives association rules based on support, confidence, and lift metrics.
Machine Learning Unsupervised Learning 5
Which problem can arise when using K-Means with very high-dimensional data?