Data Science Methods Flashcards

Question 1

What is the first step in a typical data science project?

Accepted Answer

The first step is defining the question clearly, including what decision the analysis should support and how success will be measured. Without a precise question, even technically sound work can produce results nobody can use. Stating the goal early also helps decide which data are needed and which methods are appropriate.

Question 2

Why is data cleaning considered an important stage of analysis?

Accepted Answer

Data cleaning fixes errors, duplicates, missing values, and inconsistent formats before analysis begins. Raw data often come from sensors, forms, or logs that contain mistakes, so results built on them can be misleading. Careful cleaning also documents every change, which makes the work reproducible and easier to audit later.

Question 3

What is the difference between structured and unstructured data?

Accepted Answer

Structured data are organized into rows and columns with defined fields, such as a spreadsheet or relational table. Unstructured data, such as free text, images, or audio, have no predefined schema and must be converted into usable features before standard methods can be applied.

Question 4

What does it mean to split data into training and test sets?

Accepted Answer

Splitting data means setting aside one portion for fitting a model and another portion for checking its performance on examples it has never seen. The training set teaches the model patterns while the test set estimates how well those patterns generalize. Keeping the test set untouched during development prevents an optimistic view of accuracy.

Question 5

Why should you avoid data leakage when building a model?

Accepted Answer

Data leakage occurs when information from outside the training data, especially from the test set or from the future, influences the model. It makes evaluation results look far better than the model will perform in practice. Common causes include scaling or imputing with statistics computed on the full dataset before splitting.

Question 6

What is reproducibility and why does it matter in data science?

Accepted Answer

Reproducibility means that someone else can rerun the same steps on the same data and obtain the same results. It matters because analyses inform real decisions, and errors or changes are hard to detect without a clear record. Practices such as version control, fixed random seeds, and documented code support it.

Question 7

What is a feature in a dataset, and how does it differ from a target?

Accepted Answer

A feature is an input variable, or column, that describes each observation and is used to make predictions. The target is the outcome the model tries to predict, such as house price or whether a customer will leave. Choosing and engineering good features can matter more than the choice of algorithm.

Question 8

What is the purpose of a data pipeline in a data science project?

Accepted Answer

A data pipeline is a sequence of steps that moves data from its source through cleaning, transformation, and analysis to a final output. Pipelines make work repeatable, so new data can be processed the same way without redoing each step by hand. They also make it easier to locate and fix problems.

Question 9

What do the mean and median each measure, and when is each preferred?

Accepted Answer

The mean is the arithmetic average, calculated by summing all values and dividing by the count, while the median is the middle value when data are ordered. The mean is sensitive to extreme values, so the median is usually preferred for skewed data such as household incomes.

Question 10

What is the standard deviation and what does it tell you?

Accepted Answer

The standard deviation measures how spread out values are around the mean, expressed in the same units as the data. A small value means observations cluster tightly around the average, while a large value means they vary widely. Two datasets can share a mean yet differ greatly in spread.

Question 11

What is a histogram used for during exploratory data analysis?

Accepted Answer

A histogram groups numeric values into bins and displays how often each bin occurs, revealing the shape of a distribution. It shows whether data are symmetric, skewed, or multimodal, and it can expose outliers or unexpected gaps. Choosing bin width carefully matters, because very wide or narrow bins can hide or invent patterns.

Question 12

What is correlation, and why is it not proof of causation?

Accepted Answer

Correlation is a numerical measure, ranging from negative one to positive one, of how strongly two variables move together in a linear way. Correlation cannot show that one variable causes the other, because a third factor may drive both or the relationship may be coincidental. Causation needs controlled experiments or careful study design.

Question 13

How do missing values affect analysis, and what are common ways to handle them?

Accepted Answer

Missing values can bias results and reduce statistical power if they are ignored or dropped carelessly. Common approaches include removing affected rows, imputing values with the mean, median, or a model-based estimate, and flagging missingness as its own variable. The right choice depends on why the data are missing, which should be investigated first.

Question 14

What is a box plot and what information does it display?

Accepted Answer

A box plot summarizes a distribution using its median, the first and third quartiles forming the box, and whiskers that extend toward the extreme values. Points beyond the whiskers are marked as potential outliers. Box plots make it easy to compare groups side by side and to spot differences in center, spread, and skew.

Question 15

What is sampling bias and why does it threaten conclusions?

Accepted Answer

Sampling bias occurs when the sample used for analysis systematically differs from the population it is meant to represent. Conclusions drawn from a biased sample can be wrong even when the calculations are perfect, because the data never mirrored the population. Random sampling and checking who is missing from the data reduce this risk.

Question 16

What is a confidence interval and how should it be interpreted?

Accepted Answer

A confidence interval is a range of plausible values for an unknown quantity, built by a method that captures the true value in a share of repeated samples. A 95 percent interval means the method succeeds about 95 percent of the time. It does not give a 95 percent chance for one specific interval.

Question 17

How does linear regression predict a numeric outcome?

Accepted Answer

Linear regression predicts a numeric outcome as a weighted sum of input features plus an intercept, with weights chosen to minimize the squared differences between predictions and observed values. The coefficients indicate how much the prediction changes when one feature changes, holding others constant. It assumes a linear relationship and may miss curved patterns.

Question 18

What is logistic regression, and what kind of problem does it solve?

Accepted Answer

Logistic regression is a classification method that estimates the probability an observation belongs to a class, usually by passing a weighted sum of features through a sigmoid function. It is used for yes or no outcomes, such as whether a transaction is fraudulent. Predicted probabilities are converted to class labels using a chosen threshold.

Question 19

How does a decision tree make predictions from input features?

Accepted Answer

A decision tree makes predictions by asking a sequence of yes or no questions about feature values until it reaches a leaf that holds the prediction. Splits are chosen to make resulting groups as pure as possible. Trees are easy to interpret, but a single deep tree tends to overfit noisy training data.

Question 20

What is k-nearest neighbors and how does it classify a new observation?

Accepted Answer

K-nearest neighbors classifies a new observation by finding the k closest training examples and assigning the most common label among them. The choice of k balances flexibility and smoothness: small values react strongly to noise, while large values can blur real boundaries. Because it relies on distance, features should be scaled before use.

Question 21

What is a support vector machine trying to do?

Accepted Answer

A support vector machine seeks the boundary that separates classes with the widest possible margin, using only the training points closest to that boundary, called support vectors, to define it. For data that are not linearly separable, kernel functions implicitly map features into higher dimensions where a separating boundary may exist.

Question 22

Why do ensemble methods like random forests often outperform single models?

Accepted Answer

Ensemble methods combine many models so their individual errors partly cancel out, which usually reduces variance and improves accuracy. A random forest trains many decision trees on bootstrap samples of the data and random subsets of features, then averages their predictions for regression or takes a majority vote for classification.

Question 23

What does it mean for a model to overfit, and how can you reduce it?

Accepted Answer

Overfitting means a model has learned noise and quirks of the training data, so it performs well on that data but poorly on new examples. Common remedies include gathering more data, simplifying the model, using regularization that penalizes large coefficients, and stopping training early. Cross-validation helps reveal overfitting before deployment.

Question 24

What is Naive Bayes and when is it useful?

Accepted Answer

Naive Bayes is a probabilistic classifier that applies Bayes' theorem while assuming features are conditionally independent given the class. Although that assumption is rarely exactly true, the method often works well for text classification, such as spam filtering, and it trains quickly on large sparse datasets.

Question 25

What distinguishes unsupervised learning from supervised learning in terms of labels?

Accepted Answer

Unsupervised learning works with data that have no target labels, seeking structure such as groups, patterns, or compact representations. Supervised learning uses labeled examples to learn a mapping from inputs to known outcomes. Because no correct answer is provided, unsupervised results usually require judgment and validation with domain knowledge.

Question 26

What is k-means clustering and how does it work?

Accepted Answer

K-means clustering partitions observations into k groups by repeatedly assigning each point to its nearest centroid and recalculating each centroid as the mean of its assigned points. It stops when assignments stop changing. Results depend on the starting centroids and the chosen k, so runs are usually repeated with new starting points.

Question 27

How does hierarchical clustering differ from k-means in its output?

Accepted Answer

Hierarchical clustering builds a tree of nested groups, called a dendrogram, by merging the closest clusters step by step. Unlike k-means, it does not require choosing k in advance, since the tree can be cut at any level. It is slower on large datasets because it needs pairwise distances among points.

Question 28

What is principal component analysis and what does it accomplish?

Accepted Answer

Principal component analysis is a technique that rotates data onto new axes, called principal components, ordered by how much variance each captures. Keeping only the first few components reduces dimensionality while keeping most of the variance. It assumes linear relationships, and components can be harder to interpret than the original variables.

Question 29

Why is scaling features important before applying distance-based methods?

Accepted Answer

Methods such as k-means and k-nearest neighbors treat every feature as contributing to distance, so variables with large units can dominate. Standardizing features by subtracting the mean and dividing by the standard deviation puts variables on comparable scales. Without scaling, the result may reflect measurement units instead of real structure.

Question 30

What is anomaly detection and where is it applied?

Accepted Answer

Anomaly detection identifies observations that differ markedly from the typical pattern in the data. It is applied to tasks such as fraud screening, equipment fault monitoring, and network intrusion alerts. Because anomalies are rare and often unlabeled, methods usually model normal behavior and flag points that fall far from it.

Question 31

What is association rule mining, and what is an example?

Accepted Answer

Association rule mining finds patterns in which items frequently occur together, expressed as rules such as customers who buy bread also buy butter. Its key measures include support, which is how often an itemset appears, and confidence, which is how often the consequent appears when the antecedent does. Strong rules still require human review.

Question 32

What is the curse of dimensionality and why does it matter?

Accepted Answer

The curse of dimensionality describes how data become sparse and distances become less informative as the number of features grows. With many features, a fixed amount of data covers the space poorly, which makes nearest neighbor searches and clustering less reliable. This is one reason dimensionality reduction is often used before modeling.

Question 33

Why is accuracy alone often a misleading measure of model performance?

Accepted Answer

Accuracy is the share of predictions that are correct, which can look impressive even when a model ignores the rare class of interest. If only one in a hundred transactions is fraudulent, a model that never flags fraud still scores ninety-nine percent accuracy. Precision, recall, and related metrics reveal these failures more clearly.

Question 34

What is a confusion matrix and what can you read from it?

Accepted Answer

A confusion matrix is a table that counts predictions against actual outcomes, showing true positives, false positives, true negatives, and false negatives. From it you can compute precision, recall, specificity, and accuracy. Examining the off-diagonal cells reveals which kinds of errors the model makes most often.

Question 35

How do precision and recall differ, and when does each matter most?

Accepted Answer

Precision is the share of positive predictions that are truly positive, while recall is the share of actual positives that the model finds. Precision matters most when false alarms are costly, such as flagging legitimate email as spam. Recall matters most when missing a positive is costly, such as overlooking a serious illness.

Question 36

What is cross-validation and why is it useful?

Accepted Answer

Cross-validation splits the data into several folds, trains the model on all but one fold, and evaluates it on the held-out fold, repeating until every fold has served as the test set. Averaging the results gives a more stable estimate of performance than a single split. It is valuable when data are limited.

Question 37

Why should hyperparameters be tuned using validation data rather than test data?

Accepted Answer

Hyperparameters are settings chosen before training, such as tree depth or regularization strength, and tuning them repeatedly against the test set lets information leak into the model choice. A separate validation set or cross-validation guides these choices. The test set is used once, to report an honest estimate of performance on new data.

Question 38

What is the ROC curve and what does the area under it measure?

Accepted Answer

The ROC curve plots the true positive rate against the false positive rate across all thresholds. The area under this curve summarizes how well the model ranks positive cases above negative ones, with a value of one half indicating chance-level performance. It is a common tool for comparing classifiers before choosing a decision threshold.

Question 39

What is the difference between regression metrics like mean squared error and mean absolute error?

Accepted Answer

Mean squared error averages the squared differences between predictions and actual values, so large errors are penalized heavily. Mean absolute error averages the absolute differences, treating all errors in proportion to their size. Lower values are better for both, but squared error is in squared units, so the absolute version is easier to interpret.

Question 40

What is a learning curve and what can it reveal?

Accepted Answer

A learning curve plots model performance against the amount of training data, showing training and validation scores. If both scores are low and close together, the model may be underfitting, while a large gap between them suggests overfitting. Curves still rising at the largest sample size indicate that more data may help.

Data Science Methods Flashcards

1Foundations of the Data Science Workflow

2Descriptive Statistics and Exploration

3Supervised Learning Methods

4Unsupervised Learning and Dimensionality Reduction

5Model Evaluation and Validation

Learning Resources