Unit 2: Nature of Data and Machine Learning - Subjective Questions

Data is a collection of facts, observations, measurements, images, sounds, text, or other information. In artificial intelligence, data provides the examples from which systems identify patterns and make decisions.

Importance of data:

• It helps an AI system learn about a specific problem.
• It provides examples for identifying patterns and relationships.
• It is used to train, test, and improve machine learning models.
• The quality and relevance of data strongly affect the accuracy of an AI system.

For example, a spam filter needs examples of spam and legitimate emails to learn how to classify new messages.

Data can exist in several forms:

• Text data: Emails, documents, social media posts, and reviews.
• Numerical data: Prices, ages, temperatures, and sales figures.
• Image data: Photographs, scanned documents, and medical images.
• Audio data: Speech recordings, music, and sound signals.
• Video data: Security footage, films, and online videos.
• Structured data: Information organized in tables, such as customer records.
• Unstructured data: Information without a fixed tabular format, such as photographs or natural-language conversations.

AI systems can process one or more of these forms to perform tasks such as recognition, prediction, recommendation, and classification.

Machines learn from data by examining examples and discovering useful patterns. Instead of receiving a separate instruction for every possible situation, the system is given data related to the task.

The general process is:

• Relevant data is collected.
• The data is prepared and organized.
• The machine learning system studies examples and identifies patterns.
• The system creates a model representing what it has learned.
• The model is tested using new data.
• The model is improved when its predictions are incorrect.

For example, a spam filter can study many emails labeled as spam or legitimate. It may learn that certain words, senders, or message patterns are commonly associated with spam. It can then use those patterns to classify future emails.

Machine Learning (ML) is a branch of artificial intelligence in which computer systems learn patterns from data and use those patterns to make predictions, classifications, or decisions.

Machine learning is important because it allows systems to handle tasks that would be difficult to describe with a fixed set of rules. The system improves its performance by learning from examples and experience.

Examples include:

• Classifying emails as spam or legitimate.
• Recommending products to customers.
• Recognizing objects in images.
• Predicting whether a transaction may be fraudulent.

ML is one method used to create AI systems, but artificial intelligence is the broader field concerned with making machines perform tasks associated with human intelligence.

The development of a machine learning system generally involves the following stages:

1. Define the problem: Clearly identify what the system must predict, classify, or recommend.
2. Collect data: Gather examples that are relevant to the problem.
3. Prepare the data: Remove errors, organize information, and handle incomplete or inconsistent records.
4. Train the system: Allow the machine learning method to study the prepared examples.
5. Test the system: Check how well the learned model performs on new examples.
6. Evaluate the results: Measure whether the system is accurate, reliable, and suitable for its purpose.
7. Improve and monitor: Update the system when new data or changing conditions affect its performance.

The success of the system depends not only on the learning method but also on the quality and suitability of the data.

Training data consists of examples used by a machine learning system to learn patterns. The system studies this data while building its model.

Testing data consists of separate examples used to evaluate how well the learned model works on information it has not previously seen.

Key differences:

• Training data is used for learning, while testing data is used for evaluation.
• Training examples influence the model, while testing examples should not influence its training.
• Testing helps determine whether the system can generalize beyond the examples used during learning.

Keeping the two types of data separate is important. If a system is tested only on data it has already seen, its performance may appear better than it really is.

The quality of data is important because a machine learning system learns from the examples it receives. If the data is incomplete, inaccurate, irrelevant, or biased, the system may learn incorrect patterns.

Poor-quality data can contain:

• Incorrect or outdated information.
• Missing values.
• Duplicate records.
• Irrelevant examples.
• Unbalanced representation of different groups.
• Labels that are inconsistent or incorrect.

High-quality data should be:

• Relevant to the problem.
• Accurate and consistent.
• Sufficiently varied.
• Representative of real situations.
• Properly organized and labeled when labels are required.

A machine learning model cannot reliably overcome serious problems in its training data. This idea is often summarized as "garbage in, garbage out."

A spam filter uses machine learning to identify patterns associated with unwanted messages.

Learning process:

• The system receives many examples of emails labeled spam and legitimate.
• It examines features such as words, phrases, sender information, links, formatting, and attachments.
• It identifies patterns that commonly occur in spam messages.
• It builds a model for classifying new emails.
• When a new email arrives, the model estimates whether it is likely to be spam.
• The classification can improve when users correct mistakes and provide additional examples.

A good filter must balance two errors: incorrectly allowing spam into the inbox and incorrectly sending a legitimate email to the spam folder.

A false positive occurs when a legitimate email is incorrectly classified as spam. This can cause users to miss important messages, such as work communications or account notifications.

A false negative occurs when a spam email is incorrectly classified as legitimate. This allows unwanted or potentially dangerous messages to reach the user's inbox.

Effects:

• Too many false positives make the filter inconvenient and unreliable.
• Too many false negatives reduce protection from unwanted or harmful email.
• The acceptable balance depends on the purpose of the system and user preferences.

Spam filters can be improved by using better training data, allowing users to report mistakes, and regularly updating the model for new spam patterns.

An online shopping recommendation system studies information about products and customer behavior to suggest items that may be useful or interesting.

It may use:

• Products previously viewed or purchased.
• Search queries and browsing activity.
• Ratings, reviews, and likes.
• Items frequently bought together.
• Similar behavior from other customers.
• Product categories, prices, and descriptions.

The system identifies patterns and produces recommendations for a particular customer. For example, customers who buy a camera may often purchase a memory card, so the system may recommend one.

Recommendations can change as the customer interacts with the website. However, the system must also consider privacy, inaccurate data, and the risk of repeatedly showing only similar products.

Content-based recommendation suggests products that are similar to items a customer has previously viewed, liked, or purchased. It focuses on the characteristics of the products.

Collaborative recommendation suggests products by comparing the customer's behavior with the behavior of other customers. If people with similar interests purchased a particular product, that product may be recommended.

Comparison:

• Content-based methods focus on product features and the individual user's history.
• Collaborative methods focus on patterns shared by groups of users.
• Content-based methods may work well when product descriptions are available.
• Collaborative methods may discover unexpected products but can struggle when little user data exists.

Many modern systems combine both approaches to produce more useful recommendations.

Artificial Intelligence (AI) is the broad field of creating systems that perform tasks associated with human intelligence, such as understanding language, recognizing objects, solving problems, and making decisions.

Machine Learning (ML) is a method within AI in which systems learn patterns from data instead of being given every rule directly.

Automation is the use of technology to perform a task or sequence of tasks automatically, usually according to predefined instructions. Automation does not necessarily involve learning or intelligence.

Relationship:

• ML is a subset or approach within the broader field of AI.
• Automation may be simple rule-based technology or may include AI and ML.
• A fixed program that sends scheduled emails is automation but not necessarily AI.
• A system that learns to recognize spam is ML and therefore an example of AI.

Rule-based automation follows instructions written in advance. For example, an email system may automatically forward every message from a particular address to a specified folder. The action is based on a fixed rule.

Machine learning identifies patterns from examples. For example, a spam filter may learn that certain combinations of words, links, and sender behaviors are associated with spam, even when no single rule describes every case.

Main differences:

• Automation follows predefined instructions.
• Machine learning discovers patterns from data.
• Automation usually behaves predictably when the rules remain unchanged.
• Machine learning can adapt when it is retrained with new examples.
• Both can be combined in one system.

Thus, automatic operation alone does not prove that a system is intelligent or uses machine learning.

In traditional programming, a programmer writes explicit rules that process input data and produce an output. The program's behavior depends mainly on the instructions written by the programmer.

In machine learning, the system is given examples of data and, in many cases, desired outcomes. It uses these examples to learn a model that can produce outputs for new data.

Traditional programming:

• Rules are written directly by people.
• The behavior is based on predefined instructions.
• It is useful when the rules are clear and stable.

Machine learning:

• Patterns are learned from examples.
• It is useful when the rules are complex or difficult to describe.
• Its performance depends strongly on the data used for learning.

For example, writing every rule for recognizing spam may be difficult, while a machine learning system can learn from labeled emails.

A large amount of data does not automatically guarantee correct predictions. Errors may occur for several reasons:

• Poor data quality: The examples may contain errors, missing information, or incorrect labels.
• Bias: The data may not represent all users or situations fairly.
• Irrelevant information: The system may learn patterns that do not actually matter.
• Changing conditions: Real-world behavior may change after the system has been trained.
• Unusual cases: New situations may differ from the examples seen during training.
• Overfitting: The system may memorize training examples instead of learning general patterns.
• Ambiguous examples: Even people may disagree about the correct classification.

Therefore, systems should be tested carefully using representative new data and monitored after deployment.

Learning patterns from data means identifying regular relationships, characteristics, or repeated behaviors in examples. The system does not simply memorize one answer; it uses observed patterns to make decisions about new cases.

Example 1: Spam filtering
A system may notice that certain combinations of suspicious links, phrases, and sender behaviors frequently occur in unwanted emails.

Example 2: Shopping recommendations
A system may notice that customers who purchase a particular product often view or purchase another related product.

The learned patterns help the system make predictions or recommendations. However, patterns are not always meaningful, so the system must be evaluated to determine whether they work reliably in real situations.

Feedback provides information about whether a system's output was useful or incorrect. It helps developers and models respond to changing patterns and user preferences.

In spam filtering:

• Users can mark an unwanted message as spam.
• Users can recover a legitimate message incorrectly placed in the spam folder.
• These corrections provide new examples for improving classification.

In recommendation systems:

• Users may click, rate, purchase, ignore, or hide recommended products.
• These actions indicate which suggestions are more relevant.
• The system can update future recommendations based on the new behavior.

Feedback should be collected responsibly because it may be incomplete, accidental, or biased. Regular monitoring is also needed to ensure that improvements do not create new problems.

A label is the known answer or category attached to a training example. In supervised machine learning, labels guide the system by showing what output is expected for each example.

For spam detection, emails may be labeled:

• Spam for unwanted or suspicious messages.
• Legitimate for normal messages that should reach the inbox.

The system studies the relationship between email characteristics and these labels. It then uses the learned relationship to classify new, unlabeled emails.

Labels must be accurate and consistent. If many emails are labeled incorrectly, the system may learn the wrong patterns. Creating reliable labels can require human review and clear classification guidelines.

Benefits:

• Recommendations can be personalized for individual customers.
• Systems can process large amounts of product and customer data.
• Customers may discover products that match their interests.
• Recommendations can change as customer behavior changes.
• Businesses may improve customer satisfaction and product visibility.

Limitations:

• New users may receive poor recommendations because little data is available.
• The system may recommend products based on incomplete or inaccurate behavior data.
• Personalization can create privacy concerns.
• Customers may repeatedly see similar products and miss alternatives.
• Biased data may unfairly promote some products or reduce the visibility of others.

A responsible system should provide useful recommendations while protecting privacy and allowing users appropriate control.

Relevant data contains information that helps the system solve the intended problem. If the data is unrelated, the system may waste resources or learn misleading relationships.

For example, a spam filter needs email content, sender information, and message patterns. Information about unrelated events, such as a user's favorite color, is unlikely to help decide whether an email is spam.

Relevant data helps because:

• It focuses learning on meaningful characteristics.
• It can improve the accuracy of predictions.
• It reduces unnecessary complexity.
• It makes evaluation more closely reflect the real task.

However, relevant data must also be accurate, representative, and collected responsibly. Relevance alone is not sufficient if the information is biased or incomplete.