Unit 2: Data Science Methodology

A spreadsheet with predefined columns for name, class, and attendance stores records in a consistent row-and-column format. This makes it structured data. Audio recordings, video clips, and free-text comments generally do not follow a fixed tabular structure, so they are usually treated as unstructured data.

Machine learning enables a computer system to learn patterns or relationships from example data and use them to produce outputs or predictions for new data. Therefore, option A is correct. Option D describes manually programmed, rule-based software rather than the central idea of machine learning.

Training data provide examples from which a machine-learning model learns patterns and relationships by adjusting its internal parameters. Therefore, their primary purpose is to train the model. Validation data are commonly used for hyperparameter selection, while separate test data are used for final performance evaluation.

Test data are used after model training to estimate how well the trained model performs on previously unseen examples. They should not normally be used to train the model or repeatedly tune its hyperparameters, because doing so can make the evaluation less reliable. Therefore, option A is correct.

The governing concept is supervised learning, in which a model learns from labelled examples. Each training input is paired with a known correct output, allowing the model to discover a relationship and use it for predictions on unseen data. For example, emails may be paired with spam or not-spam labels. Unlabelled data belongs mainly to unsupervised learning, while fixed rules alone are not supervised training examples.

In unsupervised learning, the data is not supplied with predefined target labels. The model examines similarities, differences, and relationships to discover patterns or groups, with clustering being a common example. Predicting labeled targets is generally a supervised-learning task.

In classification, a model uses input features to assign an example to one of several predefined classes. For example, labeling an email as spam or not spam is classification. Predicting a continuous number is regression, while grouping points and reducing features are different tasks.

In regression, a model estimates a numerical value for a target variable, such as house price, temperature, or sales amount. The result is often continuous. Class labels and yes-or-no outcomes are generally associated with classification rather than ordinary regression tasks.

In artificial intelligence, a model is a mathematical or computational representation of patterns, relationships, or rules learned from data. It uses this learned information to make predictions or decisions for new inputs. Hardware, storage systems, and user interfaces may support an AI system, but they are not the model itself.

A label specifies the correct class or target outcome associated with a data example. In supervised learning, a model learns the relationship between input features and their labels. A feature is an input property, whereas a label is the expected output. Cleaning data and splitting a dataset are separate activities, not the purpose of a label.

Data cleaning involves identifying and correcting, removing, or appropriately handling erroneous entries, inconsistent formats, duplicate records, and missing values. These steps improve data quality and make analysis more reliable. Permanently deleting a dataset, stopping model training, or randomly changing values without inspection is not the purpose of data cleaning.

A missing value means that a value in one or more data fields of a record has not been recorded, is unavailable, or is represented by a blank or special placeholder. Therefore, option A is correct. Duplicate records and incorrect units are separate data-quality problems, while option B describes complete data rather than missing data.

Overfitting occurs when a model learns the training data, including its specific details or noise, too closely. It may achieve very good training performance but perform poorly on unseen data. Option B describes good generalization, while option C more closely describes underfitting.

Underfitting occurs when a model is too simple, has unsuitable features, or is not trained sufficiently to capture the important structure in the data. It generally produces relatively high error on both training and unseen test data and is associated with high bias. Memorizing training examples too closely describes overfitting, good performance on new data describes generalization, and missing values are a data-quality issue rather than underfitting itself.

A confusion matrix is an evaluation table that compares actual class labels with the classes predicted by a classifier. Its cells count true positives, true negatives, false positives, and false negatives, allowing a teacher or analyst to inspect both correct decisions and types of error. It can support measures such as accuracy, precision, and recall, but it does not create training examples, reduce features, or fill missing values.

In binary classification, a true positive occurs when the model predicts the positive class and the ground-truth class is also positive. For example, a system correctly flags an email that really is spam. A positive prediction for a negative case is a false positive; a negative prediction for a positive case is a false negative; and a correct negative prediction is a true negative. Therefore, A is unambiguous.

In binary classification, a false positive is an error in which the model predicts the positive class even though the actual class is negative. For example, a system may flag a harmless email as spam. Option B is a true positive because both results are positive, option C is a true negative, and option D is a false negative. Therefore, A is correct.

Problem Scoping is the first stage of an AI project cycle. It identifies the problem, stakeholders, requirements, scope, and criteria for success before technical work begins. Data exploration, model building, and model evaluation are later activities. Clearly defining the problem first helps ensure that the collected data and developed model address the intended need.

Data collection obtains relevant and reliable information about the problem being studied. The collected data can support analysis, help identify patterns, and provide material for training or testing an AI model. Hiding a problem, changing results, shutting down an untested model, or deleting data is not the purpose of data collection.

A problem canvas helps define and understand a problem in a structured way. It can organize information about affected people, stakeholders, causes, effects, needs, and possible constraints. Data cleaning, programming, and advertising are different activities.

A stakeholder is any person, group, or organization that can affect a project, is affected by it, or has an interest in its decisions and outcomes. In an AI project, stakeholders may include users, customers, data providers, developers, managers, regulators, sponsors, and communities. Their needs and concerns can shape the problem definition and success criteria. A programmer may be one stakeholder, but is not the only one; software itself is not a stakeholder. Therefore, A is correct.

Ethical data collection requires people to receive clear information about what will be collected, why it is needed, how it will be used, and where appropriate, what risks are involved. Informed consent should be obtained when required, while personal information should be minimized, secured, and handled according to applicable rules. Options B and C remove transparency, and D falsifies evidence. Therefore, A expresses the essential ethical practice.

Good data should be accurate enough to represent the facts or observations being studied and relevant to the question or decision for which they are collected. Other useful qualities can include completeness, consistency, timeliness, validity, and reliability. Inaccurate, deliberately false, incomplete, or poorly sourced data can produce misleading analysis and weak AI models. Thus, option A identifies two important characteristics of good data.

Source reliability is the characteristic most directly connected with confidence that the supplied data are accurate, consistent, credible, and produced through dependable methods. Relevance asks whether the data fit the purpose, timeliness asks how current the data are, and cost describes the resources needed for collection. These qualities matter, but they do not by themselves establish trustworthiness. Therefore, option A is the best answer.

Clarifying the problem scope defines the project’s objectives, boundaries, target users, and requirements. It helps the team select suitable data and develop an appropriate solution. Therefore, option A is correct; deleting data, avoiding planning, or hiding results is unrelated to problem scoping.