Question 1 1 / 1 point

According to the slides, Machine Learning draws its strength from several established fields. Select all
that apply.
Statistics/AI
Philosophy and Ethics
Database Systems
Machine Learning/Pattern Recognition

Feedback
The slide explicitly identifies Statistics/AI as a core component, providing the theories of learning and
prediction.

The slide highlights Database Systems as crucial for providing the

technology to store, manage, and access data efficiently.

This is listed as the component that provides the actual algorithms

used to find patterns in data.

Question 2 1 / 1 point
Which of the following best describes the Knowledge Discovery in Databases (KDD) process?
A process used exclusively for visualising data and creating business reports
A multi-step process that refines raw data into actionable knowledge through stages like cleaning,
selection, and pattern evaluation
A single-step process focused only on applying a machine learning algorithm to data
A framework that deals only with storing and retrieving data from a data warehouse

Feedback
The "KDD Process View" slide explicitly outlines these multiple stages, starting from raw data in
databases and ending with final, actionable knowledge.

Question 1 1 / 1 point
According to the slides, which properties does an Ordinal attribute possess?

Distinctness, Order and Addition

 Distinctness, Order, Addition and Multiplication
Only Distinctness
Distinctness and Order

Feedback
Ordinal attributes allow you to distinguish between values (distinctness) and also place them in a
meaningful order or rank.

Question 2 1 / 1 point
What is the key difference between an Interval attribute and a Ratio attribute?
Ratio attributes have a true, meaningful zero point, while Interval attributes do not.
You can perform addition and subtraction on Ratio attributes, but not on Interval attributes.
Interval attributes can only be measured in integers, while Ratio attributes can be decimals.
Interval attributes are qualitative, while Ratio attributes are quantitative.

Feedback
The presence of a true zero is the fundamental distinction, allowing for meaningful ratio calculations
(e.g., "twice as long" is a valid comparison for a Ratio attribute but not for an Interval one like
temperature in Celsius).

Question 1 1 / 1 point
A university database stores student profiles in a table where each row is a student and each column is a
fixed characteristic like Student ID, Name and Major. Which primary data representation does this best
describe?
Ordered Data
Record Data
Transaction Data
Graph Data
Hide question 1 feedback

Feedback
This is a classic example of Record Data, where each record (a student profile) consists of a fixed set of
attributes.

Question 2 1 / 1 point
Which of the following is presented in the slides as a prime example of Graph Data?
A map showing temperature changes over time
The World Wide Web, with web pages as nodes and hyperlinks as edges
A table of customer transactions in a grocery store
A sequence of genetic nucleotides (A, C, G, T)
Hide question 2 feedback

Feedback
The slides explicitly use HTML links to illustrate Graph Data, where the relationships (links) between
objects (pages) are key.

Question 3 1 / 1 point
In the context of Document Data, what does a 'term vector' represent?
A summary of the document's main topics
A list of all the sentences contained within a document
A graphical representation of how different documents are linked together
A document where each distinct term from the entire collection is an attribute and the attribute's
value is the term's frequency in that document
Hide question 3 feedback

Feedback
Each document is converted into a vector where the components correspond to the counts of each term.

Question 1 1 / 1 point
What does the principle "Garbage In, Garbage Out" imply in the context of data science?
As long as the algorithm is powerful, the quality of the input data does not matter.
Using low-quality, unreliable data for analysis will lead to low-quality, unreliable results and
models.
Data should be discarded if it is not in a perfectly clean format.
It is more important to have a large quantity of data than high-quality data.
Hide question 1 feedback

Feedback
This is the essence of the principle; the quality of the output is directly dependent on the quality of the
input.

Question 2 1 / 1 point
According to the slides, which of the following is a reason for missing values in a dataset?
The data values are too large to be stored.
The attribute may not be applicable to all cases, such as 'annual income' for a child.
The data was intentionally duplicated to increase the dataset size.
The data contains too much random error or variance.
Hide question 2 feedback

Feedback
The slides explicitly mention that attributes may not be applicable to all data objects, leading to
legitimately missing values.

Question 3 1 / 1 point
What is the crucial distinction between 'noise' and an 'outlier' as described in the slides?
Noise is an extreme value, while an outlier is a random error.
Both noise and outliers are random errors in the data.
Noise only affects numeric data, while outliers can affect both numeric and categorical data.
Noise is considered a random error or modification of a true value, whereas an outlier is a
legitimate but considerably different data object.
Hide question 3 feedback

Feedback
The slides define noise as an error (e.g., a data entry mistake) and an outlier as a real but extreme value
(e.g., a CEO's salary in an income dataset).

Question 1 1 / 1 point
What is the main reason to perform normalisation on a dataset?
To fill in missing values using the mean or median of an attribute
To merge data from multiple sources into a single dataset
To reduce the number of attributes by selecting only the most relevant ones
To prevent attributes with large values from dominating the results in distance-based calculations
Hide question 1 feedback

Feedback
The slides highlight that normalisation scales data to a common range, ensuring that attributes like
'Income' (large values) don't unfairly outweigh attributes like 'Age' (smaller values).

Question 2 1 / 1 point
Which data preprocessing task would be primarily responsible for resolving the "Entity Identification
Problem," where customer_id in one table and cust-id in another table refer to the same person?
Data Reduction
Data Transformation
Data Cleaning
Data Integration
Hide question 2 feedback

Feedback
Data Integration is the task of merging data from multiple sources and a key challenge within it is
identifying and consolidating records that refer to the same real-world entity.

Question 1 1 / 1 point
A marketing team has a large dataset of customer purchase histories but no predefined customer
groups. They want to identify natural segments of customers (e.g., "budget shoppers," "tech
enthusiasts") to target with specific campaigns. Which machine learning paradigm is most appropriate
for this task?
Unsupervised Learning
Reinforcement Learning
Regression
Supervised Learning
Hide question 1 feedback

Feedback
Unsupervised Learning is used to find hidden patterns and structures in unlabelled data, such as
identifying natural clusters of customers.

Question 2 1 / 1 point
A real estate company wants to build a model that predicts the exact sale price of a house (e.g.,
$452,500) based on features like its size, number of bedrooms and location. What type of supervised
learning problem is this?
Clustering
 Regression
Classification
Unsupervised Learning
Hide question 2 feedback

Feedback
Regression is the task of predicting a continuous numerical value, such as the exact price of a house.

Question 1 1 / 1 point
In the two-step process of building a classification model, what is the primary purpose of the "Model
Testing" phase?
To collect and preprocess the raw data for model building
To learn the patterns and relationships from the labelled data
To select the most appropriate learning algorithm for the task
To estimate the model's accuracy on new, unseen data and check for overfitting
Hide question 1 feedback

Feedback
The testing phase uses an independent test set to evaluate how well the trained model generalises to
new data it has never seen before, which helps in estimating its real-world accuracy.

Question 2 1 / 1 point
What is the most critical rule regarding the training set and the test set?
Both sets should contain identical data to ensure consistency.
The test set must be independent of the training set.
The test set should be larger than the training set.
The test set should not contain any class labels.
Hide question 2 feedback

Feedback
If the model is tested on data it has already seen during training, the performance estimate will be
overly optimistic. This is referred to as data leakage or "cheating" and prevents a true assessment of the
model's ability to generalise.

Question 3 1 / 1 point
Which of the following are examples of a classification task? Select all that apply.
 Categorising a customer email as "support ticket," "invoice question," or "general feedback."
Diagnosing whether a tumour is malignant or benign based on an MRI scan
Determining if a credit card transaction is fraudulent or legitimate
Forecasting the total sales revenue for the next quarter
Hide question 3 feedback

Feedback
This is a multi-class classification problem, where the model must assign the email to one of three or
more predefined categories.

This is another example of binary classification, predicting one of two possible medical outcomes.

This is a binary classification problem, where the output is one of two distinct categories.

According to the error curve shown in the slides, what does it signify when both the training error and
the test error are high and close together on the left side of the graph?
Overfitting
The presence of noise in the data.
Underfitting
The model is a "Good Fit"
Hide question 1 feedback

Feedback
When the model is too simple, it fails to capture the underlying patterns in the data, resulting in poor
performance (high error) on both the data it was trained on and new data.

Question 1 1 / 1 point
A data scientist trains a classification model and observes that it has a 99% accuracy on the training
data but only 65% accuracy on the test data. What is the most likely problem affecting this model?
Lack of data
Overfitting
Low interpretability
Underfitting
Hide question 1 feedback

Feedback
A large gap between high training accuracy and low test accuracy is the classic symptom of overfitting,
where the model has "memorised" the training data, including its noise, and fails to generalise to new
data.

Question 1 1 / 1 point
In the "CC Fraud Detection" application, what serves as the 'class attribute' that the model learns to
predict?
The date and time of the transaction
Whether a transaction was labelled as {fraud, fair}
The location where the credit card was used
The total amount of money spent in the transaction
Hide question 1 feedback

Feedback
The "Application 1: CC Fraud Detection" slide explicitly states that the class attribute is the label
assigned to past transactions, indicating whether they were fraudulent or legitimate.

Question 2 1 / 1 point
In a typical classification workflow, the model's performance and accuracy are evaluated using the
same training set that was used to build it. Is this statement true or false?
True
False
Hide question 2 feedback

Feedback
The "Training vs. Testing" slide explains that a separate Test Set is used to evaluate the accuracy of the
model. The model is built on the training set and then validated on unseen data from the test set.

Question 3 1 / 1 point
What is the primary difference between Classification and Regression?
Classification is used for numerical data, while Regression is used for text data.
Classification predicts a discrete category or class, while Regression predicts a continuous,
numerical value.
Classification is a form of Unsupervised Learning, while Regression is a form of Supervised
Learning.
There is no significant difference; the terms are used interchangeably.
Hide question 3 feedback
 Feedback
The slides "The Two Tasks of Supervised Learning" and "Classification vs. Regression" explicitly state
that classification predicts a category (e.g., dog or cat) and regression predicts a numerical value (e.g.,
vehicle average, price).

Question 1 1 / 1 point
Which of the following were mentioned as real-world applications of Association Rule Discovery?
Select all that apply.
Inventory Management
Marketing and Sales Promotion
Supermarket Shelf Management
Document Clustering
Hide question 1 feedback

Feedback
This application was described for an appliance repair company wanting to stock its vehicles with parts
and tools that are frequently required together for repairs.

The slides explain how a rule like {Samosa} --> {Coke} can be used to create promotions, manage
sales, and bundle products.

This was cited as a classic application, using point-of-sale data to optimise the placement of items that
are frequently bought together.

Question 2 1 / 1 point
What is the primary goal of a clustering algorithm as described in the slides?
To label unlabeled data for use in Supervised Learning
To minimise the distance between points within a cluster, while maximising the distance between
different clusters
To ensure all data points are as close to each other as possible
To find a specific, pre-defined pattern that a user is searching for
Hide question 2 feedback

Feedback
The "Illustrating Clustering" slide clearly states that the algorithm aims to achieve two things: minimise
intra-cluster distances and maximise inter-cluster distances.

Question 1 1 / 1 point
Detecting a highly unusual credit card transaction that deviates from a user's normal spending habits is
a primary application of which machine learning task?
Classification
Clustering
Regression
Anomaly Detection
Hide question 1 feedback

Feedback
This was presented in the Lesson 4 slides as a key task for "finding the odd one out" or "detecting
significant deviations from normal behaviour," with credit card fraud being a prime example.

Question 2 1 / 1 point
According to the "Curse of Dimensionality," adding more features or attributes to a dataset always
makes it easier for machine learning algorithms to find meaningful patterns. Is this statement true or
false?
True
False
Hide question 2 feedback

Feedback
The "Challenge 2: Dimensionality" slide explains that when the number of features is very high, the
data becomes "sparse," which makes it much harder for ML algorithms to find meaningful patterns.

Question 3 1 / 1 point
The principle of "Garbage In, Garbage Out" is most directly related to which Machine Learning
challenge?
Dimensionality
Privacy Preservation
Data Quality
Scalability
Hide question 3 feedback

Feedback
The "Challenge 4: Data Quality" slide emphasises that a model's quality is entirely dependent on the
quality of the input data, highlighting issues like noise, missing values, and inconsistencies. This
concept is commonly referred to as "Garbage In, Garbage Out."

Question 1 1 / 1 point
Which phase of the decision tree process involves using a labelled training set to build the model?
Deduction (Applying)
Classification
Pruning
Induction (Learning)
Hide question 1 feedback

Feedback
Induction is the training phase where the algorithm learns from data with known outcomes to construct
the decision tree model.

Question 2 1 / 1 point
In the anatomy of a decision tree, which component represents the final predicted class label?
Branch (or Edge)
Root Node
Leaf Node (or Terminal Node)
Decision Node
Hide question 2 feedback

Feedback
The Leaf Node is the endpoint of a branch and provides the final classification outcome or prediction.

Question 1 1 / 1 point
What is the most common method for handling a Continuous attribute like 'Taxable Income' in a
decision tree?
Find an optimal threshold value 'v' and create a binary split (e.g., Income < v and Income >= v)
Convert the attribute to a nominal type by ignoring the numeric values
Always split the data into three bins: Low, Medium and High
Create a separate branch for every unique value of the attribute
Hide question 1 feedback

Feedback
Finding the best binary split by identifying an optimal cut-off point is a standard and effective method
for continuous attributes.

Question 2 0 / 1 point
Which of the following are valid ways to split a Nominal attribute like "Car Type: {Family, Sports,
Luxury}"? Select all that apply.
A binary split based on a condition like "Car Type < Sports"
A binary split grouping {Family, Sports} vs. {Luxury}
A binary split grouping {Family, Luxury} vs. {Sports, Family}
A multi-way split with one branch for each value (Family, Sports, Luxury)
Hide question 2 feedback

Feedback
The subsets in a binary split must be non-overlapping. This is not a valid grouping because "Family" is
in both proposed subsets.

A multi-way split is a straightforward method for nominal attributes.

Question 3 1 / 1 point
Decision tree algorithms are described as using a "greedy" strategy. What does this mean?
The algorithm never stops splitting until all leaf nodes are 100% pure.
The algorithm makes the locally optimal choice at each node, without considering if this choice
will lead to a globally optimal tree.
The algorithm attempts to find the single best tree by looking at all possible combinations of
splits.
The algorithm uses as much data as possible at each split to ensure accuracy.
Hide question 3 feedback

Feedback
The greedy strategy optimises at each step by selecting the split that looks best at that moment, without
backtracking or looking ahead.

Question 1 1 / 1 point
What is the primary goal of the algorithm when choosing a split, in terms of node impurity?
To create child nodes that are purer (more homogeneous) than the parent node
To create child nodes that have the same level of impurity as the parent node
To create child nodes that are more impure (less homogeneous) than the parent node
To make the child nodes as large as possible
Hide question 1 feedback

Feedback
The "best" split is one that reduces impurity, resulting in child nodes that are more homogeneous,
where one class dominates.

Question 2 1 / 1 point
If a node contains 10 records, with 5 belonging to Class A and 5 to Class B, what would its Gini Index
be?
0
0.5
0.25
1
Hide question 2 feedback

Feedback
The calculation is GINI = 1 - [ (5/10)² + (5/10)² ] = 1 - [ 0.25 + 0.25 ] = 0.5. This represents the
maximum impurity for a two-class problem.

Question 3 1 / 1 point
How is "Information Gain" used by decision tree algorithms like ID3?
It measures the reduction in entropy achieved by a split and the algorithm picks the split with the
highest gain.
It measures how much a node's classification error is reduced.
It measures the total entropy of the parent node.
It is another name for the Gini Index.
Hide question 3 feedback

Feedback
Information Gain is defined as Entropy(Parent) - Weighted Average Entropy(Children). The algorithm's
goal is to maximise this value at each split.
Question 1 1 / 1 point
According to the video “From Induction to Application: Pruning, Overfitting and Advantages”, which
of the following are key advantages of using decision trees? Select all that apply.
They are relatively inexpensive to construct and fast at classifying new records.
They require no data pre-processing at all.
They are "white box" models and easy to interpret.
They always provide the highest possible accuracy compared to all other algorithms.
Hide question 1 feedback

Feedback
The greedy construction is computationally efficient and the tree traversal for prediction is very fast.

Interpretability is a primary advantage, as the decision logic is visible and easy to explain to
stakeholders.

Question 2 1 / 1 point
What is the primary characteristic of an overfit decision tree?
It is too simple and performs poorly on both the training and test sets.
It performs equally well on the training and test sets, regardless of its complexity.
It is very small and has only a few leaf nodes.
It performs extremely well on the training set but poorly on new, unseen data (the test set).
Hide question 2 feedback

Feedback
Overfitting occurs when the model learns the noise and specific quirks of the training data, failing to
generalise to new data. This results in low training error but high test error.

Question 1 1 / 1 point
A newly generated rule is found to have very high accuracy but very low coverage. How would you
best interpret this?
The rule is very general and applies to most of the dataset, but it is often wrong.
The rule is not useful because its accuracy and coverage are both poor.
The rule is both very general and very reliable, making it an ideal rule.
The rule is very specific, reliable when it applies, but is not triggered very often.
Hide question 1 feedback
 Feedback
High accuracy means the rule is reliable. Low coverage means it only applies to a small, specific subset
of the data.

Question 2 1 / 1 point
Which of the following best describes the core idea of a rule-based classifier?
It uses complex mathematical formulas to find a hyperplane that separates data into different
classes.
It uses a collection of "IF-THEN" statements to classify records based on whether they satisfy a
set of conditions.
It classifies new data by finding the most similar data points in the training set and taking a vote.
It creates a tree-like model of decisions where each internal node represents a test on an attribute.
Hide question 2 feedback

Feedback
As stated in the video "Introduction to Rule-Based Classification" , a rule-based classifier uses a
collection of simple "IF-THEN" rules to make predictions in a human-readable way.

Question 3 1 / 1 point
Based on the "AllElectronics" dataset and the rule R1: IF age = youth THEN buys_computer = yes,
what is the Accuracy of the rule?
40%
14.28%
35.71%
100%
Hide question 3 feedback

Feedback
The rule covers 5 records where age=youth. Out of these 5, it correctly predicts 'buys_computer=yes'
for only 2 of them (RID 9 and 11). Therefore, the accuracy is 2/5 = 40%.

What is the "Indirect Method" for generating classification rules?

Learning rules one by one directly from the data using an algorithm like RIPPER
Manually writing rules based on expert domain knowledge
First building a classification model like a decision tree and then extract the rules from that model
Starting with a simplified rule and making it more complex by adding conditions until it is 100%
 accurate
Hide question 1 feedback

Feedback
The Indirect Method involves using a different model (like a decision tree) as an intermediate step to
generate an initial set of rules.

Question 2 1 / 1 point
What is a primary consequence of simplifying a rule set that was originally extracted from a decision
tree?
The accuracy of the rules on the training data always increases.
The rules become more complex and harder to interpret.
The rule set may no longer be mutually exclusive or exhaustive.
The total number of rules in the rule set increases.
Hide question 2 feedback

Feedback
Simplification (pruning conditions) makes rules more general, which can cause a single record to
trigger multiple rules (loss of exclusivity) or no rules at all (loss of exhaustiveness).

Question 3 1 / 1 point
When an initial set of rules is extracted from a decision tree without any simplification, it has two
special properties. Select all that apply.
Exhaustive
Mutually Exclusive
Simplified
Ordered
Hide question 3 feedback

Feedback
A decision tree is designed to classify any record, so every possible data record is covered by at least
one rule from the extracted set.

Because any data record can only follow one path from the root to a leaf, no two rules can be triggered
for the same data record.

Question 1 1 / 1 point
The "Learn One Rule" subroutine, as used in direct methods, typically follows what kind of search?
An exhaustive search, where it evaluates every single possible rule to find the absolute best one.
A random search, where it tries different combinations of conditions at random to find a good
rule.
A general-to-specific search, starting with the most general rule and greedily adding conditions.
A specific-to-general search, starting with a very specific rule and making it more general.
Hide question 1 feedback

Feedback
The process starts with a very general rule (e.g., IF () THEN class=positive) and iteratively adds the
condition that most improves a quality measure like information gain.

Question 2 1 / 1 point
According to the video “Advanced Rule-Based Systems & Direct Rule Mining”, what is a primary
advantage of rule-based classifiers, especially an ordered list of IF-THEN rules?
They are not affected by the scale of the data and require no pre-processing.
They are highly interpretable and human-readable.
They require no training time since the rules are predefined.
They are always more accurate than decision trees.
Hide question 2 feedback

Feedback
The video emphasises that one of the biggest advantages is that experts can easily inspect and validate
the logic of an IF-THEN rule set.

Question 3 1 / 1 point
How does a decision list (or ordered rule set) resolve the conflict that occurs when a data record
triggers two rules with different predictions?
It calculates the accuracy of both triggered rules and chooses the prediction from the rule with
higher accuracy.
It assigns the prediction of the first rule it encounters in its pre-defined order and ignores all
subsequent rules.
It uses a voting scheme where the prediction with the most rules supporting it wins.
It triggers a default rule that predicts the majority class of the entire dataset.
Hide question 3 feedback

Feedback
A decision list has a prioritised order and the first rule that "fires" for a given record makes the final
decision.

Question 4 1 / 1 point
The Sequential Covering algorithm attempts to find all the best rules for a class simultaneously by
analysing the entire dataset at once. Is this statement true or false?
True
False
Hide question 4 feedback

Feedback
The Sequential Covering algorithm is an iterative process that learns one rule at a time, removes the
positive examples covered by that rule, and then repeats the process on the remaining data.

Question 1 (1.5 points)

Which action occurs immediately after a rule is learned in the Sequential Covering algorithm?

Question 1 options:

The rule is simplified using post-pruning.

Covered positive examples are removed from the training set.

All remaining examples are relabeled.

The algorithm terminates.

Question 2 (1.5 points)

Which of the following best explains why Decision Trees are considered suitable for interpretability-sensitive
domains such as finance or healthcare?

Question 2 options:

They always outperform other supervised learning models in accuracy.

Their decision-making process can be clearly traced and explained step by step.

They require very little training data compared to other models.

They do not suffer from overfitting under any circumstances.

 Question 3 (1.5 points)

Why is a greedy search strategy used in Learn-One-Rule instead of an exhaustive search?

Question 3 options:

Exhaustive search guarantees poorer accuracy.

Exhaustive search is computationally infeasible for realistic datasets.

Greedy search avoids conflicts automatically

Greedy search guarantees optimal global rules.

Question 4 (1.5 points)

Why are rules extracted from a decision tree initially guaranteed to be mutually exclusive?

Question 4 options:

They are ordered in a decision list.

Each rule corresponds to a unique root-to-leaf path in the tree.

Each rule uses a different class label.

Each rule is simplified before extraction.

Question 5 (1.5 points)

Which property ensures that every possible data record will trigger at least one rule?

Question 5 options:

Mutually Exclusive

Exhaustive

Ordered

Simplified
Question 6 (1.5 points)

In a decision list, why is the order of rules critical?

Question 6 options:
 Because rules are evaluated in parallel.

Because later rules override earlier rules.

Because the first matching rule determines the final prediction.

Because the default rule is evaluated first.

Question 7 (1.5 points)

Which pruning technique involves first growing a full tree and then removing branches that do not improve
validation performance?

Question 7 options:

Pre-pruning

Post-pruning

Random pruning

Induction
Question 8 (1.5 points)

Why must the inherent ordering of an ordinal attribute be preserved during a split?

Question 8 options:

Because ordinal attributes cannot be split into more than two branches.

Because ordinal attributes must be converted into nominal attributes first

Because violating the order results in illogical and meaningless partitions.

Because impurity measures cannot be calculated otherwise.

Question 9 (1.5 points)

Which statement best explains why Coverage alone is insufficient to evaluate the usefulness of a rule?

Question 9 options:

A rule with high coverage always has high predictive accuracy.

A rule may cover many records but still be incorrect for most of them.

Coverage already includes information about misclassified records.

 Coverage can only be computed for simplified rules.
Question 10 (1.5 points)

True or False: A very deep decision tree that achieves near-zero training error is always desirable.

Question 10 options:
True
False
Here are the correct answers:
Question 1:
✅ Covered positive examples are removed from the training set.
(Sequential Covering removes examples already explained by the learned rule.)

Question 2:
✅ Their decision-making process can be clearly traced and explained step by step.
(Decision Trees are highly interpretable because each path shows the reasoning.)

Question 3:
✅ Exhaustive search is computationally infeasible for realistic datasets.
(Greedy search is used because checking all possible rules would take too much time.)

Question 4:
✅ Each rule corresponds to a unique root-to-leaf path in the tree.
(This makes the rules mutually exclusive.)

Question 5:
✅ Exhaustive
(Exhaustive rules ensure every possible record is covered by at least one rule.)

Question 6:
✅ Because the first matching rule determines the final prediction.
(In decision lists, rules are evaluated from top to bottom.)

Question 7:
✅ Post-pruning
(The tree is first fully grown and then unnecessary branches are removed.)
Question 8:
✅ Because violating the order results in illogical and meaningless partitions.
(Ordinal attributes have a natural order that must be preserved.)

Question 9:
✅ A rule may cover many records but still be incorrect for most of them.
(Coverage does not guarantee accuracy.)

Question 10:
✅ False
(A very deep tree may overfit the training data.)
✅ Overall summary of the report:
• No stroke
• No bleeding
• No tumor
=> Some moderate small vessel changes
Age-related brain changes
AICA vascular loop on the left side
These findings are usually managed with medicines and lifestyle control.

✅ Important things doctors usually recommend:

• Control blood pressure
• Control diabetes
• Control cholesterol
• Regular walking / exercise
• Healthy diet