WorksheetsWeb Mining 2.3
Total questions: 121
Worksheet time: 1hrs 1mins
Name
Class
Date
1.
In the slides, what is clustering described as?
a)
Organizing data elements into groups with similar properties
b)
Ordering data elements by time of arrival
c)
Assigning labels using a trained classifier
d)
Compressing data by removing duplicates
e)
Encrypting data for secure storage
2.
In clustering, what is true about elements within the same cluster?
a)
They are similar to each other and different from elements in other clusters
b)
They must have identical attribute values
c)
They must all share the same class label provided by a teacher
d)
They are chosen to maximize randomness
e)
They are always evenly spaced in the feature space
3.
Which is an application of clustering mentioned in the slides?
a)
Clustering customer groups based on interests to design marketing strategies
b)
Sorting web pages by URL length
c)
Converting images to grayscale
d)
Compiling a program into machine code
e)
Encrypting customer passwords
4.
In k-means, what does a centroid represent in the algorithm description?
a)
A cluster center that represents a cluster
b)
A single outlier point
c)
A decision tree split
d)
A class prior probability
e)
A random noise variable
5.
Which step happens after assigning each point to its nearest centroid in k-means?
a)
Recalculate centroids based on current clusters
b)
Delete the farthest points
c)
Merge the two closest clusters
d)
Normalize all attributes again
e)
Increase k automatically
6.
According to the slides, which is a stopping criterion for k-means?
a)
The sum of squares of error is less than a threshold
b)
All points become identical
c)
The data becomes linearly separable
d)
The dendrogram reaches a fixed height
e)
The classifier reaches 100 percent accuracy
7.
In the slides, what does SSE refer to in k-means convergence conditions?
a)
Sum of squares of error
b)
Standard score estimate
c)
Single-step error
d)
Similarity score expansion
e)
Squared set entropy
8.
In the slides, what is mj in k-means notation?
a)
The centroid (average vector) of cluster j
b)
The jth data point in the dataset
c)
A random threshold for stopping
d)
The distance between two clusters
e)
The maximum linkage value
9.
In k-means, what does dist(x, mj) denote in the slides?
a)
Distance between data point x and centroid mj
b)
Distance between two centroids only
c)
Probability that x belongs to cluster j
d)
Number of points in cluster j
e)
A normalization constant
10.
The k-means procedure shown in the slides alternates between which two main operations?
a)
Cluster assignment and centroid recomputation
b)
Feature selection and pruning
c)
Model training and model testing
d)
Graph construction and graph cutting
e)
Sampling and encryption
11.
In disk-k-means, what is sj used for during an iteration?
a)
An accumulated sum vector of points assigned to cluster j
b)
A set of labels for cluster j
c)
A distance threshold for cluster j
d)
A list of nearest neighbors for cluster j
e)
A noise penalty term
12.
In disk-k-means, what does nj represent?
a)
The number of points assigned to cluster j
b)
The number of clusters in the model
c)
The number of attributes in each point
d)
The number of iterations executed
e)
The number of outliers removed
13.
In disk-k-means, how is the centroid mj updated after scanning the dataset?
a)
By dividing the sum vector sj by the count nj
b)
By taking the farthest point from the cluster
c)
By averaging all centroids across clusters
d)
By selecting a random point from the dataset
e)
By minimizing a dendrogram cut value
14.
What time complexity is stated for k-means in the slides (in terms of t, k, n)?
a)
O(t k n)
b)
O(n log n)
c)
O(k log n)
d)
O(t n^2)
e)
O(k^2 n)
15.
According to the slides, k-means is only applicable to data where what exists?
a)
A mean (average) exists
b)
A median exists
c)
A categorical label exists
d)
A graph exists
e)
A time stamp exists
16.
For discrete data, which alternative is suggested in the slides instead of k-means?
a)
k-modes algorithm
b)
k-nearest neighbors
c)
Apriori algorithm
d)
PageRank
e)
Naive Bayes
17.
What does the slide state about k-means and outliers?
a)
It is sensitive to outliers
b)
It is immune to outliers
c)
It removes outliers automatically
d)
It requires outliers to converge
e)
It only works when outliers are many
18.
What does the slide state about k-means initialization?
a)
It is sensitive to initialization and may approach local optima
b)
It always finds the global optimum regardless of initialization
c)
Initialization is unnecessary because centroids are fixed
d)
Initialization only matters for hierarchical clustering
e)
Initialization is only required for classification
19.
According to the slides, k-means is not suitable for clusters with what shape property?
a)
Non-spherical (not super-sphere) shapes
b)
Perfectly spherical shapes
c)
One-dimensional line shapes
d)
Exactly two clusters
e)
Clusters with equal sizes
20.
The outlier example in the slides is used to highlight which issue for k-means?
a)
Outliers can lead to unwanted clustering results
b)
k-means cannot compute any distances
c)
Outliers always form their own correct cluster
d)
Outliers are required for centroid updates
e)
Outliers eliminate the need for normalization
21.
The random initialization illustration is used to emphasize that k-means results can change mainly due to what?
a)
Different initial centroid choices
b)
Different file formats
c)
Different dendrogram levels
d)
Different graph cut algorithms
e)
Different TF-IDF vocabularies
22.
The example with two natural non-spherical clusters is used to show that k-means can fail because it assumes clusters are best represented by what?
a)
Centroids of roughly spherical or elliptical regions
b)
Decision boundaries of a neural network
c)
Ordered rules on discrete values only
d)
Graph partitions based on mincut
e)
A single root node in a tree
23.
Which cluster representation in the slides is described as fitting elliptical or spherical clusters?
a)
Centroid-based representation
b)
Rule-based representation
c)
Graph-cut representation
d)
Suffix-tree representation
e)
Hash-based representation
24.
In the slides, what is assumed in the 'based on the classification model' representation of clusters?
a)
Each cluster corresponds to a class label
b)
Each cluster must contain exactly k points
c)
Each cluster must be a single connected component in a graph
d)
Each cluster is defined by a fixed threshold t
e)
Each cluster is defined only by cosine similarity
25.
Which representation is said to fit discrete values, including text?
a)
Based on common values in the cluster
b)
Centroid-based representation
c)
Ward method representation
d)
SVM margin representation
e)
Euclidean ball representation
26.
In hierarchical clustering, what structure is used to represent nested clusters?
a)
A tree structure called a dendrogram
b)
A confusion matrix
c)
A neural network
d)
A hash table
e)
A FIFO queue
27.
In a dendrogram, what do the leaves represent according to the slides?
a)
Individual data points
b)
Only cluster centroids
c)
Only class labels
d)
Only outliers
e)
Only features
28.
What is the key idea of bottom-up (agglomerative) clustering described in the slides?
a)
Repeatedly merge the closest pair of clusters until one remains
b)
Start with one cluster and split it into two at each step
c)
Randomly assign points to clusters
d)
Optimize a margin using labeled data only
e)
Build a classifier from positive documents only
29.
What is the key idea of top-down (divisive) clustering described in the slides?
a)
Start with all data in one cluster and recursively split into subclusters
b)
Merge the two farthest clusters at each step
c)
Select k random centroids and iterate
d)
Convert all attributes to binary and sum distances
e)
Use TF-IDF and cosine similarity to label data
30.
In the agglomerative algorithm described, how are clusters initialized at the start?
a)
Each data point is treated as its own cluster
b)
All data points are placed into one cluster
c)
k centroids are selected randomly
d)
Only labeled points form initial clusters
e)
Clusters are initialized using TF-IDF vectors only
31.
In agglomerative clustering, what happens after combining the two closest clusters into a new cluster c?
a)
Compute the distance from c to the other clusters
b)
Immediately stop the algorithm
c)
Remove all remaining clusters
d)
Normalize the dataset again
e)
Assign class labels to every point
32.
In single-link hierarchical clustering, the distance between two clusters is defined as what?
a)
The shortest distance between any pair of points across the two clusters
b)
The maximum distance between any pair of points across the two clusters
c)
The average distance between cluster centroids only
d)
The increase in SSE after merging
e)
The distance between the two most central points
33.
What issue is specifically mentioned for single-link clustering in the slides?
a)
It can cause chaining effects due to noise
b)
It cannot handle continuous attributes
c)
It requires labeled data
d)
It always produces balanced clusters
e)
It cannot be implemented on disk
34.
In full-link hierarchical clustering, the distance between two clusters is defined as what?
a)
The maximum distance between any pair of points across the two clusters
b)
The shortest distance between any pair of points across the two clusters
c)
The average distance between all pairs of points across the two clusters
d)
The distance between two random points
e)
The cosine similarity between centroids
35.
What is stated about full-link clustering compared to single-link?
a)
It avoids chaining effects but is sensitive to outliers
b)
It causes more chaining effects but is robust to outliers
c)
It always runs in linear time
d)
It requires min-max normalization
e)
It only works with binary attributes
36.
In medium-link clustering, the distance between two clusters is defined as what?
a)
The average distance between points across the two clusters
b)
The minimum distance between points across the two clusters
c)
The maximum distance between points across the two clusters
d)
The centroid-to-centroid distance only
e)
The increase in classification error
37.
In the Ward method mentioned in the slides, the distance between two clusters is based on what?
a)
The increase in the sum of squared errors when the clusters are combined
b)
The shortest point-to-point distance across clusters
c)
The maximum point-to-point distance across clusters
d)
The number of shared labels in the clusters
e)
The Jaccard similarity of binary vectors
38.
What advantage of hierarchical clustering is stated in the slides?
a)
It allows generating clusters by choosing a level of the tree
b)
It always finds the global optimum for k-means objective
c)
It requires no distance function
d)
It automatically determines the best feature set
e)
It guarantees linear runtime
39.
What disadvantage of hierarchical clustering is stated in the slides?
a)
High computational and storage costs
b)
It cannot be visualized as a tree
c)
It requires labeled data for training
d)
It only works with discrete attributes
e)
It cannot create nested clusters
40.
Which distance function family is listed for continuous attributes in the slides?
a)
Minkowski distance
b)
Jaccard distance
c)
Hamming code
d)
TF-IDF scoring
e)
Entropy gain
41.
Which of the following is a distance measure listed under continuous attributes in the slides?
a)
Manhattan distance
b)
Jaccard similarity
c)
Information gain
d)
Cosine-Rocchio score
e)
F-score
42.
Which distance measure listed in the slides uses attribute weights in its definition?
a)
Weighted Euclidean distance
b)
Chebychev distance
c)
Jaccard distance
d)
Single-link distance
e)
Z-score distance
43.
For symmetric binary attributes, how is the distance between two data points defined in the slides (in words)?
a)
Mismatch count divided by total number of binary attributes
b)
Number of matches divided by total attributes
c)
Sum of squared differences of numeric values
d)
Maximum difference across attributes
e)
Average of cluster centroids
44.
For asymmetric binary attributes, which distance is named in the slides?
a)
Jaccard distance
b)
Euclidean distance
c)
Chebychev distance
d)
Ward distance
e)
Cosine distance
45.
For general discrete attributes (non-binary), how is distance described in the slides (in words)?
a)
Based on the number of mismatches across attributes
b)
Based only on the maximum numeric difference
c)
Based only on the mean value of attributes
d)
Based on the Euclidean length of the vector
e)
Based on the SSE increase after merging
46.
Why do the slides recommend normalizing data before distance-based methods?
a)
To prevent attributes with large ranges from dominating the distance
b)
To increase the number of clusters automatically
c)
To eliminate the need for a distance function
d)
To guarantee zero SSE
e)
To convert supervised learning into unsupervised learning
47.
Which normalization approach is described by scaling values to a fixed interval such as 0 to 1?
a)
Min-max (range) normalization
b)
Hierarchical normalization
c)
Graph cut normalization
d)
Random projection
e)
Bagging
48.
Which normalization approach is described by subtracting the mean and dividing by standard deviation?
a)
Z-score normalization
b)
Min-max normalization
c)
Single-link normalization
d)
Rocchio normalization
e)
Spying normalization
49.
For exponential continuous attributes, what transformation is suggested in the slides?
a)
Logarithmic transform
b)
One-hot encoding
c)
Median subtraction
d)
Cluster merging
e)
Dendrogram pruning
50.
Discrete attributes with no order (for example, fruits) can be converted to what form according to the slides?
a)
Binary attributes
b)
Exponential attributes
c)
Graph edges
d)
Centroid vectors only
e)
Class priors
51.
Ordered discrete attributes (for example, age) are suggested to be handled similarly to what?
a)
Linear continuous normalization
b)
Jaccard distance computation
c)
Mincut labeling
d)
Spying thresholding
e)
Ward linkage
52.
When data has multiple attribute types, what is the first recommended strategy in the slides?
a)
Convert all attributes to the most common type (for example, linear continuous)
b)
Discard all non-numeric attributes
c)
Use only cosine similarity
d)
Train a supervised classifier instead of clustering
e)
Sort attributes by name and cluster
53.
After converting or standardizing attribute types, how do the slides propose combining distances across attributes?
a)
Compute distance per attribute and sum them
b)
Take the maximum distance across attributes only
c)
Multiply all attribute distances together
d)
Use only the first attribute distance
e)
Replace distances with class labels
54.
In the mixed-attribute distance formula, what does delta equal 1 indicate?
a)
The attribute exists in both data points being compared
b)
The attribute is missing in both points
c)
The attribute is symmetric binary
d)
The attribute is exponential
e)
The attribute is the class label
55.
In clustering evaluation, what does a user-based method rely on in the slides?
a)
A panel of experts whose ratings are averaged
b)
Only the SSE value
c)
Only the number of clusters k
d)
Only a max-flow computation
e)
Only the value of c in PU learning
56.
In clustering evaluation, what does 'based on classification' mean in the slides?
a)
Use labeled (classified) data where each class corresponds to a cluster
b)
Use only unlabeled data and graph cuts
c)
Use only expert opinions
d)
Use only compression metrics
e)
Use only isolation metrics
57.
What does the slides' compression metric measure?
a)
Concentration of points around the centroid (for example, SSE)
b)
Separation between cluster centroids
c)
Number of nodes in a dendrogram
d)
Probability that a document is positive
e)
Number of attributes shared by two points
58.
What does the slides' isolation metric measure?
a)
Degree of separation of clusters via distances between centroids
b)
How many outliers exist in the dataset
c)
How many attributes are asymmetric
d)
The likelihood of missing data
e)
The computational cost of max-flow
59.
What is indirect evaluation in clustering as described in the slides?
a)
Evaluating clustering by performance on a downstream end task
b)
Evaluating clustering only by runtime
c)
Evaluating clustering only by number of clusters
d)
Evaluating clustering only by expert ratings
e)
Evaluating clustering by dendrogram height
60.
Why do the slides say supervised learning can be difficult in some applications like web document classification?
a)
Labeling requires manual effort and labels can change over time
b)
Distances cannot be computed for text
c)
Centroids cannot be updated for text
d)
Graphs cannot be constructed for text
e)
Unlabeled data cannot be collected
61.
What is LU learning as described in the slides?
a)
Learning from a small set of labeled examples plus many unlabeled examples to improve a classifier
b)
Learning only from unlabeled data without any labels
c)
Learning only from positive examples without any unlabeled data
d)
Learning a clustering tree without distances
e)
Learning a classifier only from expert ratings
62.
What is the main purpose of the EM algorithm described in the slides?
a)
Iteratively maximize a probability estimate when data is missing
b)
Directly compute exact centroids in one pass
c)
Build a dendrogram without distances
d)
Remove outliers deterministically
e)
Guarantee global optimum for k-means
63.
In co-training, the attribute set is divided into how many sets in the slides?
a)
Two sets (X1 and X2)
b)
Three sets
c)
One set only
d)
k sets (one per cluster)
e)
n sets (one per example)
64.
What is the key data setup for PU learning described in the slides?
a)
A set P of positive documents and a set U of unlabeled documents containing positive and negative
b)
A set of labeled positives and labeled negatives only
c)
A set of clusters and a dendrogram only
d)
A set of centroids and outliers only
e)
A set of k nearest neighbors only
65.
You have two numeric attributes where one ranges from 0 to 1 and the other ranges from 0 to 1000. Before using a distance-based method, what is the best step to reduce bias from scale?
a)
Normalize or standardize the attributes to comparable scales
b)
Increase k until SSE is zero
c)
Convert both attributes to text tokens
d)
Use only the attribute with the larger range
e)
Remove all points with small values
66.
Your dataset is purely discrete and you want a k-based partitioning method similar to k-means. Which method aligns with the slide guidance?
a)
k-modes
b)
Single-link hierarchical clustering
c)
Ward method
d)
Max-flow mincut
e)
Spectral graph partitioning
67.
If k-means keeps reassigning many points between clusters across iterations, which stopping condition would NOT yet be satisfied?
a)
The number of reassigned points is below a threshold
b)
The dataset has missing labels
c)
The attribute types are mixed
d)
The graph is fully connected
e)
The vocabulary size is fixed
68.
Two runs of k-means on the same data produce different final clusters. Based on the slides, what is the most likely reason?
a)
Different initial centroid choices led to different local optima
b)
The distance function changed automatically
c)
Hierarchical clustering was applied in between
d)
The algorithm switched to k-modes mid-run
e)
Normalization forces different k values
69.
A single extreme outlier is added far from the rest of the points. For k-means, which effect is most consistent with the slide discussion?
a)
Centroids can be pulled toward the outlier, changing assignments
b)
The outlier is always ignored during centroid updates
c)
SSE becomes exactly zero
d)
k-means automatically increases k by one
e)
The outlier forces a dendrogram to be built
70.
You suspect your clusters are non-spherical and may form ring-like or elongated shapes. Why might k-means give poor results according to the slides?
a)
It represents clusters by centroids and fits spherical or elliptical regions
b)
It cannot compute any distance on continuous data
c)
It requires labeled negatives to operate
d)
It can only build a tree of nested clusters
e)
It always merges the closest pair of clusters
71.
You are using hierarchical clustering and want to avoid the chaining effect caused by noise. Which linkage choice best matches the slide trade-off?
a)
Full-link (complete linkage)
b)
Single-link linkage
c)
Centroid distance linkage
d)
Jaccard linkage
e)
Z-score linkage
72.
You also want the linkage method to be less sensitive to outliers than full-link, while not suffering from chaining like single-link. Which slide-listed option is the best compromise?
a)
Medium-link (average linkage)
b)
Single-link linkage
c)
Full-link linkage
d)
k-modes
e)
Min-max normalization
73.
If you merge two clusters and measure their distance by the increase in sum of squared errors after merging, which method are you using?
a)
Ward method
b)
Single-link method
c)
Full-link method
d)
Jaccard distance
e)
Cosine-Rocchio
74.
For binary attributes where the value 1 is more important than 0 (for example, presence of a rare feature), which distance is more appropriate per the slides?
a)
Jaccard distance (asymmetric binary)
b)
Symmetric binary distance
c)
Euclidean distance
d)
Chebychev distance
e)
Ward distance
75.
For general discrete attributes, distance increases when two data points match on fewer attributes. Which description best fits the slide definition?
a)
Distance is based on the number of mismatches across attributes
b)
Distance is the maximum numeric difference only
c)
Distance is always zero if any attribute matches
d)
Distance is the Euclidean length of the vector
e)
Distance is the SSE increase after merging clusters
76.
You have exponential-valued continuous attributes. Before computing distances, what preprocessing step is suggested in the slides?
a)
Apply a logarithmic transform
b)
Convert values to binary using one-hot encoding
c)
Replace all values with their ranks only
d)
Drop the attribute entirely
e)
Force all values into two clusters
77.
In mixed-type distance computation, an attribute is missing for one of the two points. According to the slide notation with delta, what is the correct handling?
a)
Set delta to 0 so that attribute does not contribute to the summed distance
b)
Set delta to 1 and treat missing as zero
c)
Set the distance for that attribute to the maximum possible
d)
Copy the value from the other point
e)
Always stop and discard the pair
78.
You are clustering text documents and there is no ground-truth class labeling available. Which evaluation approach best matches the slide options?
a)
User-based evaluation using expert judgments
b)
Classification-based evaluation using labeled classes
c)
Compression metric only (SSE)
d)
Isolation metric only (centroid separation)
e)
Max-flow complexity analysis
79.
Clustering is used as an intermediate step inside a recommender system pipeline. Which evaluation approach aligns with the slide definition?
a)
Indirect evaluation using end-task performance
b)
User-based evaluation only
c)
Isolation metric only
d)
Single-link chaining analysis
e)
Minkowski distance comparison
80.
If you want to measure how tightly points in each cluster concentrate around a centroid, which evaluation metric category is the best match?
a)
Compression metric
b)
Isolation metric
c)
User-based metric
d)
Indirect metric
e)
Covariate shift metric
81.
If you want to measure how separated clusters are by looking at distances between their centroids, which evaluation metric category is the best match?
a)
Isolation metric
b)
Compression metric
c)
User-based metric
d)
Indirect metric
e)
Spying metric
82.
In the hole discovery procedure, why are random points added and labeled as N?
a)
To train a classifier that learns regions with little or no original data
b)
To compute exact k-means centroids
c)
To guarantee the dendrogram is balanced
d)
To eliminate the need for normalization
e)
To ensure all clusters have equal size
83.
In LU learning, how can unlabeled data improve a classifier built from few labeled examples?
a)
It can reveal useful co-occurring features that strengthen the model
b)
It provides true class labels directly
c)
It removes the need for any labeled data
d)
It guarantees independence of features
e)
It turns clustering into classification automatically
84.
In EM-based LU learning, why does the algorithm alternate between assigning probabilistic labels and re-estimating parameters?
a)
Each step improves the model based on the other until parameters stabilize
b)
It is required to compute Euclidean distance
c)
It is the same as single-link clustering
d)
It guarantees k equals the number of classes
e)
It eliminates all outliers by construction
85.
Co-training assumes the attributes can be split into two views X1 and X2. What is a key additional assumption stated in the slides?
a)
X1 and X2 are independent with respect to the class label
b)
X1 and X2 must be identical feature sets
c)
X1 must be binary and X2 must be continuous
d)
X1 must have higher variance than X2
e)
X2 must contain all missing attributes
86.
Compared with self-training, what is a distinguishing characteristic of co-training in the slides?
a)
It uses two classifiers trained on two attribute sets
b)
It uses only one classifier and no confidence scores
c)
It requires a dendrogram to select examples
d)
It requires reliable negative examples RN
e)
It uses only k-means centroids as labels
87.
In graph-based LU methods, what is the purpose of connecting each vertex to its k nearest neighbors?
a)
To build a sparse similarity graph that captures local relationships
b)
To force the graph to be fully connected with all pairs
c)
To ensure all edges have equal weight
d)
To guarantee balanced partitions automatically
e)
To convert text into binary attributes
88.
Why do spectral graph methods normalize the cut objective as described in the slides?
a)
To discourage very unbalanced partitions that plain mincut tends to choose
b)
To make SSE equal to zero
c)
To ensure k-means converges faster
d)
To eliminate the need for edge weights
e)
To guarantee all unlabeled points become positive
89.
In PU learning, why is it necessary to build a classifier without reliable negative labels?
a)
Because only positive examples are confidently labeled while U mixes positive and negative
b)
Because distance functions do not work on text
c)
Because dendrograms cannot be built from unlabeled data
d)
Because k-means requires a mean to exist
e)
Because Jaccard distance cannot be computed
90.
In the two-step PU classifier construction, why is identifying a reliable negative set RN useful?
a)
It provides trustworthy negative training data to build a better classifier
b)
It forces all unlabeled documents to be positive
c)
It eliminates the need for any classifier
d)
It increases the number of positive examples in P
e)
It converts binary attributes into continuous ones
91.
In the spying technique, why is the probability distribution of spy documents S used to set threshold t?
a)
Because S are known positives placed in U, so their scores help separate positives from likely negatives
b)
Because S are known negatives placed in P, so their scores detect outliers
c)
Because S determine k automatically
d)
Because S provides the dendrogram height
e)
Because S removes the need for TF-IDF
92.
Cosine-Rocchio uses TF-IDF vectors and cosine similarity. What is the main reason this representation is used in the method description?
a)
It enables comparing documents to a positive-set vector using cosine similarity
b)
It guarantees Euclidean distance is minimized
c)
It converts all features into labels
d)
It ensures hierarchical clustering has low storage cost
e)
It makes k-means insensitive to outliers
93.
In the 1DNF technique, a document in U is added to RN when it satisfies which condition?
a)
It contains no words from the positive feature set PF
b)
It contains many PF words
c)
It is closest to the centroid of P
d)
It has the maximum cosine similarity to P
e)
It is labeled as positive by both classifiers
94.
In I-SVM, why are documents classified as negative repeatedly moved from Q into RN?
a)
To iteratively expand the reliable negative set and refine the SVM boundary
b)
To make k-means converge in fewer iterations
c)
To compute the Ward linkage distance faster
d)
To keep recall fixed at 1
e)
To guarantee the graph becomes fully connected
95.
In PU probability estimation, under the Totally Random Selection Hypothesis, why is f(x) computed as g(x) divided by c?
a)
Because g(x) estimates labeling probability and c is the constant selection rate among positives
b)
Because g(x) is always larger than 1 and must be scaled down
c)
Because c is the number of clusters and g(x) must be normalized
d)
Because c is the SSE threshold used for convergence
e)
Because g(x) is a distance and c converts it to similarity
96.
If you only need to rank documents by how likely they are to be positive, what does the slide note about using g(x) directly?
a)
g(x) can be used directly because monotonic scaling by c does not change ranking
b)
g(x) cannot be used because it requires negative labels
c)
g(x) must be squared to become a probability
d)
g(x) must be converted to Jaccard distance first
e)
g(x) must be replaced by SSE
97.
You want the ability to choose different numbers of clusters after running the algorithm by cutting at different levels. Which slide-described property supports this goal?
a)
Hierarchical clustering produces a tree that can be cut at different levels
b)
k-means always outputs all possible k values at once
c)
Disk-k-means stores every iteration as a dendrogram
d)
Jaccard distance automatically selects k
e)
Spying technique outputs a cluster tree
98.
You have an unordered categorical attribute (for example, fruit type) and want to include it in a distance calculation. What preprocessing is recommended in the slides?
a)
Convert it to binary attributes (one-hot style)
b)
Treat categories as continuous numbers without changes
c)
Drop all continuous attributes instead
d)
Use only Ward linkage and skip preprocessing
e)
Force the attribute to be exponential and log-transform it
99.
A dataset contains two ring-shaped clusters plus a few extreme outliers. Based on slide limitations, which combination of issues most strongly explains why k-means may produce misleading clusters?
a)
Centroid-based spherical bias plus sensitivity to outliers
b)
Need for labeled data plus high storage cost
c)
Inability to compute any distance plus chaining effect
d)
Requirement that k is unknown plus need for TF-IDF vectors
e)
Dependence on dendrogram level plus independence of views
100.
You need hierarchical clustering on noisy data, but you also expect occasional extreme outliers. Which linkage choice is most defensible using the slide trade-offs?
a)
Medium-link (average linkage)
b)
Single-link linkage
c)
Full-link linkage
d)
Jaccard distance linkage
e)
Ward method is always best regardless of data
101.
In mixed-attribute distance computation, many attributes are missing for some pairs of points. Which handling best matches the slide formula logic and avoids artificially inflating distance?
a)
Exclude missing attributes using delta = 0 so only shared attributes contribute
b)
Treat missing values as zeros and always include them
c)
Assign a fixed maximum penalty for every missing attribute
d)
Impute missing values using random sampling for each pair
e)
Stop computing distance whenever any attribute is missing
102.
If you skip normalization when one attribute has a range much larger than others, what is the most likely downstream consequence for clustering results?
a)
Assignments become driven mainly by the large-range attribute, masking other attributes
b)
The algorithm becomes immune to outliers
c)
The dendrogram becomes perfectly balanced
d)
The number of clusters k is automatically corrected
e)
The SSE necessarily decreases to zero
103.
For asymmetric binary attributes where 1 indicates presence of an important property and 0 indicates absence, why is Jaccard distance more appropriate than symmetric binary distance?
a)
It ignores joint absences and focuses on mismatches involving presences
b)
It guarantees Euclidean geometry for the data
c)
It makes all clusters spherical
d)
It enforces independence of X1 and X2
e)
It removes the need for unlabeled data
104.
In EM-based LU learning with very few labeled examples, what is a key failure mode implied by the iterative labeling and re-training loop?
a)
Early misclassifications in U can be reinforced in later parameter updates
b)
Outliers are automatically removed so clusters vanish
c)
The method cannot use probabilities for classes
d)
The algorithm always terminates after one iteration
e)
The method requires a dendrogram cut to proceed
105.
Co-training adds high-confidence predictions from f1 to train f2 and vice versa. If the two views X1 and X2 are not independent for the class label, what is the most likely risk?
a)
The two classifiers may amplify the same bias and propagate errors
b)
The algorithm becomes equivalent to k-modes
c)
The method turns into hierarchical clustering automatically
d)
The method can no longer compute cosine similarity
e)
The method guarantees balanced cuts in the graph
106.
In graph-based LU methods, what can happen if the similarity threshold is set too high when selecting edges?
a)
The graph can become too sparse or disconnected, making label propagation unreliable
b)
The graph becomes fully connected and always balanced
c)
The number of classes increases automatically
d)
All unlabeled vertices become positive regardless of similarity
e)
The maximum flow complexity drops to linear time
107.
The slides note that mincut tends to produce unbalanced sets. What is the most accurate explanation for this tendency?
a)
A small isolated set can yield a low cut weight even if sizes are very uneven
b)
Mincut requires Euclidean distances which are biased
c)
Mincut always merges the closest clusters first
d)
Mincut can only handle symmetric binary attributes
e)
Mincut ignores edge weights by definition
108.
In PU learning, the slides discuss approximating the objective by minimizing the rate of predicting positive on U while enforcing a recall constraint. What trade-off does this approximation manage?
a)
It reduces false positives on U while trying to keep most true positives found
b)
It maximizes SSE while minimizing centroid distance
c)
It guarantees perfect precision and perfect recall simultaneously
d)
It eliminates the need for any labeled positives in P
e)
It forces all unlabeled documents to be classified as negative
109.
In the spying technique, the threshold t is determined using the spy set S inserted into U. If S is not representative of P (for example, it contains unusually easy positives), what is a plausible failure outcome?
a)
The threshold can be set too high, causing many true positives in U to be treated as reliable negatives
b)
The threshold is always correct because S are labeled
c)
All unlabeled documents become positive regardless of t
d)
The method becomes identical to Ward linkage
e)
The method cannot compute any probabilities
110.
In the 1DNF technique, RN is initialized as U and then documents containing any PF word are removed from RN. If PF is too small, what is the most likely consequence?
a)
RN may become overly large and include many hidden positives as false negatives
b)
RN becomes empty and no classifier can be trained
c)
PF automatically expands until perfect separation
d)
k-means will fix the mistake by changing k
e)
The graph becomes fully connected and balanced
111.
The slides mention covariate shift in PU learning applications. Which risk does covariate shift most directly create for techniques that select negatives from U using similarity thresholds?
a)
Thresholds tuned on one distribution may misidentify negatives when the unlabeled distribution shifts
b)
Cosine similarity becomes undefined on TF-IDF vectors
c)
Ward method cannot be computed anymore
d)
Z-score normalization stops working on continuous data
e)
EM can no longer re-estimate parameters
112.
I-SVM iteratively expands RN by moving negatives from Q. If RN initially contains many hidden positives, what is a likely failure mode?
a)
The SVM boundary can be biased and may incorrectly remove additional positives from Q
b)
The algorithm will automatically relabel RN as positive
c)
The algorithm becomes equivalent to co-training with two views
d)
The method guarantees convergence to the true boundary
e)
The method cannot use slack variables
113.
The PU SVM formulation includes a case where P may contain noise (negative examples). Why introduce separate weights for positive and negative errors (C+ and C-)?
a)
To control the relative penalty of mistakes when positives may be noisy and unlabeled negatives are uncertain
b)
To make Euclidean distance smaller than Manhattan distance
c)
To enforce a dendrogram cut at a fixed level
d)
To ensure attributes X1 and X2 are independent
e)
To convert discrete attributes into continuous ones
114.
The probability estimation approach uses f(x) = g(x) / c under the Totally Random Selection Hypothesis. If c actually depends on x, what breaks most directly?
a)
The scaling by a constant no longer yields calibrated estimates of Pr(y = 1 | x)
b)
Cosine similarity can no longer be computed
c)
The k-means time complexity changes from O(t k n) to O(n)
d)
Ward method stops merging clusters
e)
The definition of a dendrogram becomes invalid
115.
c is estimated by averaging g(x) over positive validation examples. If the validation positives include some mislabeled negatives, what is the most likely impact on f(x) estimates?
a)
c can be underestimated, inflating f(x) and making positives look more likely than they are
b)
c becomes exactly 1, so f(x) equals g(x)
c)
c becomes negative, making f(x) undefined
d)
f(x) becomes independent of x by definition
e)
The method switches to Jaccard distance automatically
116.
Spectral graph methods normalize the cut by a factor involving the sizes of the two sets. Which outcome is this normalization trying to prevent?
a)
A very small partition being chosen simply because it yields a low raw cut weight
b)
Clusters with spherical shape being split by centroids
c)
Outliers being ignored during k-means updates
d)
Binary attributes being treated as continuous
e)
TF-IDF vectors being normalized twice
117.
Ward method measures distance by the increase in SSE when merging clusters. If your goal is to preserve a notion of compactness similar to k-means, why can Ward be more aligned than single-link?
a)
Because it directly penalizes merges that greatly increase within-cluster variance
b)
Because it guarantees no outliers exist
c)
Because it always yields a balanced dendrogram
d)
Because it ignores all distances except the minimum
e)
Because it only works on binary attributes
118.
The hole discovery method trains a classifier to detect regions labeled N. If too few random points are added, what is the most likely limitation of the resulting model?
a)
It may miss parts of empty regions and fail to learn a good boundary for holes
b)
It will always overfit perfectly and find all holes
c)
It will force k-means to converge faster
d)
It will automatically create reliable negatives RN
e)
It will compute Jaccard distance incorrectly
119.
EM alternates between estimating missing labels and updating parameters. If the model assumptions are violated (for example, text generation independence is poor), what is a plausible consequence?
a)
The parameter updates may converge to a poor solution even if they stabilize
b)
The algorithm becomes identical to single-link clustering
c)
The method no longer needs unlabeled data
d)
The classifier guarantees zero error probability
e)
The method cannot compute any class probabilities
120.
k-means requires a chosen k. If the true structure has more natural groups than your chosen k, what outcome is most consistent with k-means behavior?
a)
Some natural groups will be merged together because the algorithm must produce exactly k clusters
b)
The algorithm will automatically increase k to match the data
c)
The algorithm will output a dendrogram instead of clusters
d)
The algorithm will switch to Jaccard distance
e)
The algorithm will remove outliers until groups match k
121.
You need to evaluate clustering quality for text data, have no ground-truth labels, and want to judge whether clusters are meaningful for humans. Which evaluation approach best fits, and why?
a)
User-based evaluation, because expert panels can judge semantic coherence
b)
Compression metric, because SSE is always meaningful for text
c)
Isolation metric, because centroid separation guarantees interpretability
d)
Classification-based evaluation, because labels are unnecessary
e)
Mincut complexity, because runtime implies quality
100 %
