Machine Learning (ML) Unsupervised Learning Exercises

Option 3 is correct because Unsupervised Learning focuses on finding intrinsic structures, such as clusters or associations, in data that lacks ground-truth labels. Option 1 and 2 describe Supervised Learning tasks (Classification and Regression). Option 4 describes Reinforcement Learning, which is based on feedback loops rather than static data patterns.

Option 1 is correct because Clustering is a core unsupervised task used to group similar data points (images) together without needing prior labels.

Option 3 and 4 are impossible without an existing labeled dataset to train a model.

Option 2 is used for predicting continuous numerical values, which does not apply to the task of organizing or grouping images.

Option 2 is correct because Anomaly (or Outlier) Detection identifies rare items or events that raise suspicions by differing significantly from the majority of the data.

Option 1 is used to compress data.

Option 3 finds rules that describe your data, like 'people who buy X also buy Y'.

Option 4 typically involves time-series forecasting, often requiring historical labels.

Option 2 is correct because Dimensionality Reduction aims to reduce the number of random variables under consideration by obtaining a set of principal variables.

Option 1 is the opposite, adding more features.

Option 3 is a supervised ensemble technique.

Option 4 is a semi-supervised technique used to assign labels to unlabeled data points based on labeled neighbors.

Option 4 is the correct choice. Here is why:

• Dataset A: Contains a target label (Gender), which allows a model to learn a mapping from inputs to a known output (Supervised Learning).
• Dataset B: Lacks any labels. The model must find inherent structures or groups without guidance (Unsupervised Learning).
• Options 1 & 2: Are incorrect because they misidentify the presence or absence of labels as the defining factor.
• Option 3: Is incorrect as Reinforcement Learning relies on an agent/reward system, not static labeled/unlabeled datasets.

Option 1 is correct. Here is the breakdown:

• Association Rule Learning: Is designed specifically to discover interesting relations or "if-then" patterns between variables in large databases.
• Option 2: Is a dimensionality reduction technique used to simplify data, not find item associations.
• Option 3: Is a supervised learning algorithm used for classification/regression.
• Option 4: Is a supervised model used to predict continuous numerical values.

Option 4 is the correct choice. The reasoning is as follows:

• Redundancy: Highly correlated features provide the same information; removing one reduces noise and saves memory/time without losing the signal.
• Option 1: Unsupervised learning does not create labels; it finds structure.
• Option 2: Reducing dimensions actually simplifies the model, which can prevent overfitting but doesn't "increase complexity."
• Option 3: Dimensionality reduction is a tool within Unsupervised Learning, not a way to switch paradigms.

Option 2 is correct. Let's evaluate the table:

• Method X: Uses labels to predict future classes, which defines Classification (Supervised).
• Method Y: Works with unlabeled data to find groups, which defines Clustering (Unsupervised).
• Option 3: Is wrong because Method X is clearly supervised due to the "Labeled" input.
• Option 4: Is wrong because Method Y (Clustering) is specifically used when "Ground Truth" labels are unavailable.

Option 3 is correct. Evaluation is tricky because:

• No Ground Truth: In supervised learning, we have actual labels to check against. In unsupervised learning, "success" is often subjective or based on internal density/separation metrics.
• Option 1: Unsupervised models work extensively with numerical data (e.g., K-Means).
• Option 2: They often require larger datasets to find meaningful patterns.
• Option 4: They do use loss functions (like Inertia in Clustering), but these don't measure "accuracy" in the traditional sense.

Option 1 is correct. This is a fundamental concept:

• Similarity: Since there are no labels, algorithms must rely on distance or density to decide which points belong together.
• Option 2: This describes an accuracy check in Supervised Learning.
• Option 3: This is a description of cluster purity, not the "measure" used to create the clusters.
• Option 4: Refers to feature selection in regression/classification, requiring a target variable which Unsupervised Learning lacks.

Option 2 is correct. Here is the breakdown:

• Scale Dominance: If one feature ranges from 0–1 and another from 0–1000, the distance calculation will be biased toward the larger range.
• Option 1: Scaling is a data transformation and does not create labels.
• Option 3: This describes Dimensionality Reduction, not Scaling.
• Option 4: Scaling does not change the learning paradigm.

Option 1 is correct. Let's look at the logic:

• Anomaly Detection: Fraudulent transactions are usually rare and "distant" from normal behavior patterns in the feature space.
• Option 2: This is a simple database lookup, not machine learning.
• Option 3: This describes a Regression task.
• Option 4: This describes the process of creating a labeled dataset for Supervised Learning.

Option 4 is correct. Here is why this is a "curse":

• High Dimensions: As the number of features increases, the volume of the space grows so fast that the data becomes sparse, making all points seem equidistant.
• Option 1: This is a data volume issue, not a dimensionality issue.
• Option 2: High dimensionality usually makes models more complex, not simpler.
• Option 3: More features typically slow down the training process significantly.

Option 3 is correct. Based on the goals in the table:

• Technique A (PCA): Principal Component Analysis is the standard for reducing noise and projecting high-D data into 2D or 3D for visualization.
• Technique B (Clustering): Grouping instances based on similarity is the definition of clustering.
• Options 1, 2, & 4: These include supervised methods like Regression or Decision Trees which require labels.

Option 2 is correct. Here is the function of a dendrogram:

• Visualization: It is a tree-like diagram that records the sequences of merges or splits in hierarchical clustering.
• Option 1: Error rates are usually visualized with learning curves or residual plots.
• Option 3: This is the job of Feature Importance or PCA.
• Option 4: Unsupervised learning does not work with pre-defined labels.

Option 4 is correct. Here is the breakdown of unsupervised evaluation:

• Silhouette Coefficient: Values range from -1 to +1; a high value indicates that the object is well matched to its own cluster and poorly matched to neighboring clusters.
• Options 1 & 2: These are regression metrics used to measure the distance between predictions and actual labels.
• Option 3: This is a classification metric used to evaluate labeled predictions.

Option 1 is correct. This refers to Hierarchical Clustering:

• Agglomerative: It is a "bottom-up" approach where each observation starts in its own cluster, and pairs of clusters are merged as one moves up the hierarchy.
• Option 2: This describes the "Divisive" approach, which is "top-down."
• Option 3: This describes Label Propagation, a semi-supervised technique.
• Option 4: This is a description of a specific PCA transformation.

Option 3 is correct. Let's look at the "Lift" metric in Association Rules:

• Lift > 1: Indicates that the two items are dependent on each other, and the occurrence of Item A has a positive effect on the occurrence of Item B.
• Option 1: This would be indicated by low Support or Lift < 1.
• Option 2: This refers to high "Confidence" or "Support" for Item B alone.
• Option 4: A Lift value of exactly 1 implies independence.

Option 2 is correct. Here is how to interpret the Elbow Method:

• The Elbow: It plots the "Inertia" (sum of squared distances) against the number of clusters. The point where the decrease slows down (the "elbow") is considered the ideal cluster count.
• Option 1: This would be the goal of a Scree Plot in PCA.
• Option 3: Accuracy cannot be measured without labels; the Elbow Method measures internal cohesion.
• Option 4: While it might help see how outliers affect clusters, its primary goal is cluster count selection.

Option 2 is correct. This summarizes the core difference:

• Discovery: Unsupervised learning aims to explore the data and discover unknown patterns or structures within it.
• Option 1: This is the primary goal of Supervised learning.
• Option 3: While unsupervised results can lead to creating labels, it is not the primary mechanism of the algorithm itself.
• Option 4: This is a statistical process typically used for validating assumptions rather than a machine learning paradigm.