Unit 2: Data Science Methodology

An effective problem statement clearly and concisely describes the specific problem, identifies its purpose or users, and includes measurable aspects for evaluating success. The other options lack clarity, user focus, or evidence-based definition, so option A is correct.

Stakeholder identification is most useful during early problem scoping and planning because it reveals who is affected, what needs the project should address, what constraints exist, and how success should be judged. Early consultation can also expose risks, accessibility concerns, ethical issues, and differing expectations before major resources are committed. Feedback may continue throughout development, but waiting until the end reduces its value. Therefore, A is correct.

Success criteria state in advance what an acceptable project outcome should achieve and how it will be judged. They provide a basis for evaluating the model or solution against the intended goals. Thus, option A is correct, while the other options oppose proper evaluation.

Choosing sources beforehand helps ensure that the collected data is relevant to the objective, sufficiently reliable, and appropriate in quality. Poorly chosen sources may introduce bias, errors, or gaps. Therefore, option A is correct; collecting unnecessary data is not the goal.

A survey collects information by asking people questions about their opinions, experiences, preferences, or needs. Therefore, people’s opinions and experiences are suitable examples of survey data. Machine temperature, network speed, and soil moisture are usually measured directly with instruments or sensors.

Primary data are original data collected directly from the source for a specific purpose. The source may be a person, observation, experiment, or event. Surveys, interviews, and direct measurements are common examples. Data taken from an old news report are generally secondary data.

Secondary data are data previously collected by another person or organisation and later obtained from an existing source. Examples include books, government reports, research papers, and suitable online databases. Data collected directly from a person for the current purpose are primary data.

Duplicate records can cause some observations to be counted more than once. Consequently, certain values or categories may appear more frequently than they should, which can bias the analysis and lead a model to learn misleading patterns. Duplicates should be checked during data cleaning.

Data preprocessing prepares raw data for analysis or model training. It may include correcting errors, handling missing values, removing unsuitable duplicates, changing formats, and scaling or transforming values. It does not mean hiding all data or removing every feature from the dataset.

Normalization is a feature-scaling technique that transforms numerical values to a comparable range or scale. This prevents a feature with large units, such as income, from dominating a feature with smaller units, such as age, in algorithms affected by magnitude or distance. Therefore, option A is correct. Missing-value removal is a separate cleaning step, while normalization does not add rows or incorrect values.

Many statistical and machine-learning algorithms calculate distances, weights, probabilities, or mathematical transformations, so they generally require numerical input. Categorical values such as “red,” “blue,” or “online” can be represented through one-hot encoding or another suitable method. Thus, option A is correct. Encoding does not imply that category names are wrong, hidden, or deleted; it changes their representation for computation.

Classification predicts a discrete class or category, such as spam or not spam. Regression predicts a numerical quantity, such as house price or temperature. Both are commonly supervised-learning tasks and can be used with text, tabular, image, and other data.

The training set is used to learn patterns and model parameters, whereas the test set is kept unseen until evaluation. Separating them gives a more trustworthy estimate of how the model will perform on unfamiliar data. If test examples influence training, data leakage can produce an overly optimistic score. Option A is correct; separation does not stop training, erase data, or hide results.

In binary classification, a true negative occurs when the model predicts the negative class and the actual label is also negative. It is a correct prediction and occupies one cell of the confusion matrix. Option A is therefore correct. A true positive is predicted positive and actually positive, a false positive predicts positive incorrectly, and a false negative predicts negative when the actual class is positive.

Precision measures the reliability of positive predictions: precision = TP ÷ (TP + FP). It asks, “Of all cases predicted positive, how many were truly positive?” Therefore, option A is correct. Recall uses TP ÷ (TP + FN) and focuses on finding actual positives; accuracy considers all correct predictions; specificity measures correctly identified negatives. These metrics answer different evaluation questions.

Using diverse and balanced training data can provide better representation of the groups affected by an AI system and reduce the chance that one group is consistently disadvantaged. Developers should also evaluate results across groups, because balanced data alone does not guarantee fairness. Restricting data to one group, adding inaccurate information, or skipping evaluation can increase bias; therefore option A is the best answer.

A final review examines the project’s goals, methods, results, limitations, errors, and feedback. This process helps the team understand what worked and identify improvements for future projects. The other options contradict the purpose of evaluation, transparency, and continuous improvement.

Problem scoping is an early project-cycle activity that defines the problem, intended users, desired outcome, constraints, and boundaries of an AI solution. It helps the team decide what success means before collecting data or selecting a model. Choosing a complex algorithm, gathering unnecessary data, or deploying without evaluation ignores these requirements. Hence option B is correct.

A measurable goal provides a clear standard for judging whether a solution has succeeded. The results can be compared with the target using suitable indicators. Thus, option A is correct, while the other options contradict the purpose of a clear and measurable problem statement.

Data acquisition is the process of obtaining the relevant data needed for a defined problem from suitable sources. Sources may include surveys, sensors, records, databases, or approved public collections, depending on the project. Cleaning fixes errors, analysis searches for patterns, and training uses prepared data to build a model; these are later or separate activities. Therefore, option A is correct.

A representative sample reflects important characteristics of the target population, such as age, region, or relevant groups. Therefore, conclusions drawn from the sample can be generalized to the wider population more reliably. However, representativeness can reduce bias without completely eliminating every possible source of bias.

If one group is overrepresented while other groups are underrepresented, the data may not represent the full population fairly. An analysis or model based on such data may produce results that favour the overrepresented group. Balanced representation, diverse sources, and regular checks can help reduce this risk.

Data exploration is the initial examination of a dataset’s structure, data types, distributions, relationships, missing values, and unusual observations. It helps identify patterns and possible quality problems before further analysis. Using data without examining it can lead to misleading or incorrect conclusions.

Selecting an appropriate model depends mainly on the problem objective—such as classification, regression, or clustering—and on the nature and quality of the available data. The computer’s color, chair height, and file-name length do not determine whether a model is suitable.

Validation data are used during model development to compare models and tune hyperparameters. Training data are used to learn the model parameters, while a separate test set is normally reserved for the final performance evaluation. Therefore, validation data support development decisions rather than final unbiased reporting.