Data Mining Exam Answers - April 2024

Part A - Answers (10 × 2 = 20 marks)

1. What are the advantages of time series analysis?
Time series analysis helps identify trends, patterns, and seasonal variations in data
over time. It enables accurate forecasting, supports decision-making, and is useful for
anomaly detection in fields like finance, weather prediction, and inventory
management.
2. Write a note on prediction.
Prediction involves using historical data and statistical or machine learning models to
forecast future outcomes. It relies on patterns and relationships in the data, such as
trends or correlations, and is widely used in sales forecasting, risk assessment, and
resource planning.
3. What is sequence discovery?
Sequence discovery is the process of identifying frequent or significant sequences of
events or items in a dataset over time. It is used in applications like market basket
analysis, web click-stream analysis, and DNA sequence analysis.
4. Define: “Hypothesis Testing.”
Hypothesis testing is a statistical method used to make inferences about a population
based on sample data. It involves formulating a null hypothesis (H0) and an
alternative hypothesis (H1), then using statistical tests to accept or reject H0 based on
a significance level.
5. Mention the advantages of neural networks.
Neural networks excel at handling complex, non-linear data, adapt to new patterns
through learning, and are effective for tasks like image recognition, natural language
processing, and predictive modeling due to their layered architecture.
6. Show the example of activation functions in neural networks.
Examples of activation functions include:
o Sigmoid: ( f(x) = \frac{1}{1 + e^{-x}} ) (introduces non-linearity).
o ReLU: ( f(x) = \max(0, x) ) (reduces vanishing gradient problem).
o Tanh: ( f(x) = \tanh(x) ) (outputs values between -1 and 1).
7. What are the uses of ID3?
ID3 (Iterative Dichotomiser 3) is used to construct decision trees for classification
tasks. It selects the best attribute to split data based on information gain, making it
effective for categorical data analysis and rule generation.
8. Give the purpose of C4.5 in decision tree.
C4.5 improves upon ID3 by handling both continuous and categorical data, managing
missing values, and using gain ratio for attribute selection. Its purpose is to create
more accurate and robust decision trees for classification.
9. State the clustering with genetic algorithms.
Clustering with genetic algorithms involves using evolutionary techniques (e.g.,
selection, crossover, mutation) to optimize cluster assignments. It is useful for finding
global optima in complex, non-linear datasets.
10. What do you mean by divisive clustering?
Divisive clustering is a top-down approach where all data points start in one cluster,
which is recursively split into smaller clusters based on dissimilarity. It is the opposite
of agglomerative clustering and is effective for hierarchical clustering.
 11. Define: “Data Parallelism.”
Data parallelism is a technique where a large dataset is divided into smaller subsets,
and multiple processors or threads process these subsets simultaneously using the
same algorithm. It is commonly used to speed up data mining tasks on big data.
12. What are generalized association rules?
Generalized association rules extend traditional association rules by including
hierarchical or taxonomic relationships between items (e.g., “milk” and “dairy”).
They capture broader patterns, improving the applicability of rules in diverse datasets.

Part B - Answers (5 × 5 = 25 marks)

13. Summarize the major issues of data mining.
Data mining faces several challenges, including data quality issues like noise, missing
values, and inconsistencies, which can skew results. Scalability is a concern with
large datasets, requiring efficient algorithms. Privacy and security are critical, as
mining personal data raises ethical concerns. Overfitting, where models fit noise
rather than patterns, can reduce generalizability. Additionally, handling high-
dimensional data and selecting relevant features pose difficulties, while integrating
heterogeneous data sources complicates the process. Addressing these requires robust
preprocessing, advanced algorithms, and ethical frameworks.
14. What are the data mining from a database perspectives? Explain.
From a database perspective, data mining involves extracting useful patterns from
structured data stored in databases. It includes querying large datasets using SQL or
specialized tools to identify trends, associations, or anomalies. Key techniques include
association rule mining (e.g., market basket analysis), classification (e.g., decision
trees), and clustering (e.g., k-means). Databases provide efficient storage and
retrieval, but challenges like data redundancy, indexing, and query optimization must
be addressed. This perspective emphasizes integrating data mining with database
management systems for scalable, real-time analysis.
15. Distinguish between the regression and correlation.
Regression and correlation both analyze relationships between variables, but they
differ in purpose and output. Regression models the relationship, predicting a
dependent variable (e.g., sales) from an independent variable (e.g., advertising spend)
using an equation like ( y = mx + c ). Correlation measures the strength and direction
of a linear relationship (e.g., Pearson’s r, ranging from -1 to 1) without predicting
outcomes. Regression is causal and predictive, while correlation is descriptive and
non-directional. For instance, high correlation doesn’t imply causation, unlike
regression.
16. Describe the regression in statistical-based algorithms.
In statistical-based algorithms, regression predicts continuous outcomes by modeling
the relationship between variables. Linear regression fits a line (e.g., ( y = \beta_0 + \
beta_1x )) to minimize the sum of squared errors between observed and predicted
values. It assumes linearity, independence, and normal distribution of residuals.
Variants like logistic regression handle categorical outcomes (e.g., yes/no) using the
sigmoid function. These methods are foundational in data mining for forecasting,
requiring careful validation to avoid overfitting or multicollinearity.
17. Elaborate the simple approach of distance-based algorithms.
Distance-based algorithms cluster or classify data by measuring similarities using
distance metrics like Euclidean or Manhattan distance. A simple approach, such as k-
nearest neighbors (k-NN), assigns a data point to the class of its nearest neighbors
 based on distance. The process involves calculating distances, selecting k neighbors,
and applying a majority vote or average. This method is intuitive and effective for
small datasets but struggles with high dimensions (curse of dimensionality) and
requires careful choice of k and normalization of features.
18. Bring out the minimum spanning tree in partitional algorithms.
The minimum spanning tree (MST) is a concept from graph theory that can be
integrated into partitional algorithms in data mining, particularly for clustering tasks.
An MST is a subset of edges in a weighted, undirected graph that connects all vertices
(data points) with the minimum total edge weight, without forming cycles. In
partitional algorithms like k-means or PAM, MST can be used as a preprocessing step
or an alternative approach to initialize clusters or determine optimal partitions.

Role in Partitional Algorithms:

 Cluster Initialization: MST helps identify natural groupings by connecting data

points based on proximity (e.g., Euclidean distance). Edges with the highest weights
can be removed to form k clusters, where k is the desired number of partitions.
 Outlier Detection: Long edges in the MST often indicate outliers, which can be
excluded before applying partitional clustering, improving result quality.
 Hierarchical to Partitional Transition: MST can guide the transition from
hierarchical clustering to partitional clustering by cutting edges to achieve k
partitions.

Process:

1. Construct a graph where each data point is a vertex, and edges represent distances
between points.
2. Use an algorithm like Kruskal’s or Prim’s to find the MST, ensuring the total edge
weight is minimized.
3. For k partitions, remove the k-1 longest edges, splitting the MST into k subtrees, each
representing a cluster.
Example: For 5 points (A, B, C, D, E) with distances, the MST might connect A-B, B-
C, C-D, D-E. Removing the longest edge (e.g., D-E) splits it into two clusters.

Advantages:

 Ensures connectivity of all points with minimal distance, providing a robust initial
structure.
 Reduces sensitivity to initial centroid selection, a common issue in k-means.

Limitations:

 Computational cost increases with dataset size (O(E log V) for Kruskal’s, where E is
edges and V is vertices).
 Less effective in high-dimensional spaces where distance metrics become unreliable.

In partitional algorithms, MST enhances clustering by offering a geometric foundation,

though it requires careful edge-cutting strategies to align with the k-partition goal.
 19. Write down the measuring the quality of rules in association rules.
Measuring the quality of rules in association rule mining is essential to ensure the
extracted rules are meaningful and actionable. Association rules, such as “if bread,
then butter,” are evaluated using specific metrics that quantify their strength,
reliability, and usefulness. The primary measures include:

 Support: This measures the frequency of the rule’s itemset in the dataset, calculated
as the proportion of transactions containing both the antecedent and consequent (e.g.,
support = 60% if 60% of transactions include {bread, butter}). It indicates the rule’s
statistical significance.
 Confidence: This assesses the reliability of the rule, defined as the probability of the
consequent given the antecedent (e.g., confidence = 75% if 75% of bread transactions
include butter). It reflects the rule’s predictive strength.
 Lift: This compares the observed support of the rule with the expected support if
items were independent (e.g., lift > 1 indicates a positive correlation, such as 1.5
meaning the rule is 50% more likely than random). It measures the rule’s interest or
strength of association.

Application: These metrics help filter rules—high support ensures commonality, high
confidence ensures reliability, and high lift ensures relevance. However, challenges include
balancing trade-offs (e.g., high confidence with low support) and avoiding spurious
correlations, requiring domain knowledge for interpretation.

In association rule mining, these measures collectively determine rule quality, guiding
decisions in applications like market basket analysis.

Part C - Answers (3 × 10 = 30 marks)

20. Outline the basic data mining tasks in detail.
Data mining involves several core tasks aimed at extracting valuable insights from
data. The primary tasks include:

 Classification: Assigns data points to predefined categories (e.g., spam vs. not spam)
using algorithms like decision trees or neural networks. It requires labeled training
data and is widely used in fraud detection.
 Clustering: Groups similar data points into clusters without prior labels, using
methods like k-means or hierarchical clustering. It’s useful for customer segmentation
and pattern recognition.
 Association Rule Mining: Identifies relationships between variables (e.g., “if bread,
then butter”) using rules like support and confidence, as seen in market basket
analysis.
 Regression: Predicts continuous outcomes (e.g., sales figures) based on input
variables, employing linear or logistic regression techniques.
 Anomaly Detection: Spots unusual patterns or outliers (e.g., fraudulent transactions)
using statistical or distance-based methods.
 Summarization: Provides concise data representations, such as averages or trends, to
aid understanding. These tasks require preprocessing (e.g., cleaning data), feature
selection, and validation to ensure reliability, though over-reliance on automated tools
can sometimes overlook contextual nuances.
 21. Discuss the Bayes theorem in statistical perspective on data mining.
Bayes Theorem, expressed as ( P(A|B) = \frac{P(B|A) \cdot P(A)}{P(B)} ), is a
foundational statistical tool in data mining for probabilistic classification and
decision-making. Here, ( P(A|B) ) is the posterior probability of hypothesis A given
evidence B, ( P(B|A) ) is the likelihood, ( P(A) ) is the prior probability, and ( P(B) )
is the marginal probability. In data mining, it underpins Naive Bayes classifiers,
which assume independence between features to predict categories (e.g., email spam
detection). Its strength lies in handling small datasets and incorporating prior
knowledge, but the independence assumption can oversimplify real-world data,
potentially skewing results. Variants like Bayesian networks address this by modeling
dependencies, making it versatile for tasks like medical diagnosis or sentiment
analysis, though it requires careful tuning to avoid bias from inaccurate priors.
22. Illustrate the K nearest neighbors in distance-based algorithms.
The K Nearest Neighbors (k-NN) algorithm is a simple, instance-based method in
distance-based algorithms used for classification and regression. It operates by finding
the k closest data points (neighbors) to a new, unclassified point based on a distance
metric, typically Euclidean distance (( d = \sqrt{(x_2 - x_1)^2 + (y_2 - y_1)^2} )). For
classification, the majority class among the k neighbors determines the new point’s
label; for regression, it averages the neighbors’ values.

Illustration: Suppose we have a dataset with two classes (A and B) and a new point P. With
k=3, we calculate distances to all points, select the three nearest (e.g., 2 from A, 1 from B),
and assign P to class A. The choice of k is critical—small k (e.g., 1) is noise-sensitive, while
large k smooths but may include irrelevant points. Preprocessing like normalization is
essential to avoid bias from varying feature scales. k-NN’s strength lies in its simplicity and
adaptability, but it struggles with high-dimensional data (curse of dimensionality) and
requires significant memory for large datasets, making it less efficient for real-time
applications.

23. Demonstrate the PAM algorithm in partitional algorithms.

The PAM (Partitioning Around Medoids) algorithm is a robust partitional clustering
method that extends k-means by using medoids (actual data points) as cluster centers
instead of means. It aims to minimize the sum of dissimilarities between points and
their assigned medoid, typically using Euclidean distance.

Demonstration: Given a dataset with n points and k clusters, PAM starts by randomly
selecting k medoids. It iteratively swaps a medoid with a non-medoid point if the swap
reduces the total cost (sum of distances). For example, with 5 points (A, B, C, D, E) and k=2,
if A and C are initial medoids, PAM evaluates swapping C with D. If the new configuration
lowers the total distance, D replaces C. This process repeats until convergence.

PAM is more robust to noise and outliers than k-means because medoids are real data points,
not averages. However, its computational complexity (O(k(n-k)²)) makes it slower than k-
means, especially for large datasets. It’s ideal for small-to-medium datasets where outlier
resistance is key, though it requires careful initial medoid selection to avoid suboptimal
clustering.

24. Examine the large itemsets in association rules.

Large itemsets are subsets of items in a dataset that appear together frequently,
forming the foundation of association rule mining. This technique, often implemented
 via algorithms like Apriori or FP-Growth, identifies relationships between items (e.g.,
“bread and butter”) based on metrics like support, confidence, and lift. Large itemsets
are those exceeding a user-defined minimum support threshold, indicating statistical
significance.

Process: The Apriori algorithm, for instance, starts by generating frequent 1-itemsets (items
meeting minimum support). It then iteratively builds larger itemsets (2-itemsets, 3-itemsets,
etc.) by joining frequent sets and pruning those below the threshold using the Apriori
property: any subset of a frequent itemset must also be frequent. Example: In a transaction
dataset with {bread, milk, butter}, if {bread, butter} has 60% support (above the 50%
threshold), it’s a large 2-itemset. This continues until no new large itemsets are found.

Key Metrics:

 Support: Percentage of transactions containing the itemset (e.g., 60% for {bread,
butter}).
 Confidence: Probability of buying butter given bread (e.g., 75% if 75% of bread
transactions include butter).
 Lift: Ratio of observed support to expected support, indicating rule strength (e.g., lift
> 1 suggests positive correlation).

Challenges: Generating large itemsets is computationally intensive, especially with many

items or low support thresholds, leading to the “curse of cardinality.” Sparse datasets may
yield few large itemsets, while dense ones can produce an explosion of candidates, requiring
efficient pruning. Real-world applications, like market basket analysis or web usage mining,
benefit from optimizing these thresholds to balance relevance and performance.

Conclusion: Large itemsets enable actionable insights (e.g., product placement strategies),
but success depends on tuning parameters and handling scalability, making advanced
algorithms and parallel processing valuable for large datasets.