Clustering GMLE Practice Questions (Page 6)

Questions 26–30

1. 26application · medium

   A data scientist is clustering a dataset that contains overlapping clusters. They want to use a Gaussian Mixture Model. How does the Expectation-Maximization (EM) algorithm estimate the parameters of the Gaussian components?

   • AIt iteratively alternates between computing the probability of each point belonging to each component (E-step) and updating the component parameters based on those probabilities (M-step).
   • BIt performs a grid search over all possible parameter values to find the global optimum.
   • CIt randomly assigns points to clusters and then computes the mean and variance of each cluster.
   • DIt uses gradient descent to minimize the within-cluster sum of squares.
   Reveal Solution Discussion Select an answer first
2. 27application · medium

   A logistics company uses DBSCAN to group GPS coordinates of delivery stops into service areas. The data contains many isolated rural stops that are not part of any dense cluster. The current epsilon is set to 0.01 degrees and minPts to 5. The resulting clusters are too fragmented, and many stops are labeled as noise. The team wants to reduce the number of noise points while still identifying dense urban clusters. What should they do?

   • AIncrease minPts and keep epsilon the same
   • BSwitch to OPTICS with a lower minPts
   • CDecrease epsilon and increase minPts
   • DIncrease epsilon and keep minPts the same
   Reveal Solution Discussion Select an answer first
4. 28expert · hard

   A team is clustering GPS trajectories of vehicles. They want to identify common routes as clusters, but the data contains many short detours that should be considered noise. They use DBSCAN with a fixed epsilon and minPts. After running, they find that some known routes are split into multiple clusters, while others are merged. What is the most likely cause and best adjustment?

   • AThe minPts value is too high; lower it to ensure more points are considered core points.
   • BThe data needs to be standardized; apply z-score normalization to all features.
   • CThe fixed epsilon is too small for some routes and too large for others; use OPTICS instead, which can handle varying densities.
   • DThe distance metric is inappropriate; switch from Euclidean to Manhattan distance.
   Reveal Solution Discussion Select an answer first
5. 29expert · hard

   A data engineer is clustering a dataset with millions of records and hundreds of features. They need to run K-Means nightly. The current implementation takes too long and sometimes fails to converge. They also need to select the number of clusters automatically. Which approach best addresses both performance and model selection?

   • AUse DBSCAN and set epsilon based on a k-distance plot
   • BUse Mini-Batch K-Means and select k using the silhouette score on a sample
   • CUse hierarchical clustering with Ward's linkage
   • DUse standard K-Means with k=10 and rely on the elbow method
   Reveal Solution Discussion Select an answer first
6. 30expert · hard

   A data scientist is clustering a dataset with features that have different units and scales. They plan to use K-Means. They are debating whether to standardize the features (z-score) or normalize them to [0,1]. Which statement is most accurate?

   • ABoth methods are equivalent and produce the same clustering result
   • BNormalization to [0,1] is always better for clustering
   • CStandardization is generally preferred because it is less affected by outliers
   • DScaling is not necessary for K-Means
   Reveal Solution Discussion Select an answer first