Machine Learning Basics Flashcards

Question 1

What is machine learning in simple terms?

Accepted Answer

Machine learning is a branch of artificial intelligence in which a program improves its performance on a task by finding patterns in data, rather than being given every rule explicitly. The system adjusts internal parameters using examples, so its predictions or decisions become more accurate as it sees more relevant experience.

Question 2

What is the difference between supervised and unsupervised learning?

Accepted Answer

Supervised learning uses training examples that include correct labels, so the model learns to map inputs to known outputs. Unsupervised learning works with unlabeled data and looks for hidden structure, such as groups or compressed representations, without being told the right answer in advance.

Question 3

What are features and labels in a dataset?

Accepted Answer

Features are the input variables describing each example, such as a house's square footage or the words in an email. A label is the target value the model tries to predict, such as the sale price or whether the email is spam. Each example pairs its features with one label.

Question 4

What is a model in machine learning?

Accepted Answer

A model is the mathematical function or structure that maps input features to outputs, with adjustable parameters whose values are set during training. Training searches for parameter values that make the model's predictions match the data as closely as possible. After training, the fitted model is used to make predictions on new inputs.

Question 5

What is a training set, and why is it kept separate from test data?

Accepted Answer

A training set is the portion of data used to fit the model's parameters. Test data is held back until the end so it can estimate how well the model performs on examples it has never seen. Reusing training examples for evaluation gives an overly optimistic picture of real-world accuracy.

Question 6

What is a loss function and what role does it play?

Accepted Answer

A loss function is a formula that measures how far a model's predictions are from the correct answers, producing a single number for each example or batch. Training aims to reduce this value. Gradient descent is a common method that repeatedly adjusts parameters in the direction that lowers the loss.

Question 7

What is the difference between classification and regression?

Accepted Answer

Classification predicts a discrete category, such as whether a transaction is fraudulent or which species appears in a photo. Regression predicts a continuous numeric quantity, such as tomorrow's temperature or the price of a used car. Both are supervised tasks because they learn from labeled examples.

Question 8

What is clustering, and when would you use it?

Accepted Answer

Clustering is an unsupervised technique that groups examples so that items within a group are more similar to each other than to items in other groups. It is useful when you have no labels and want to explore structure, such as segmenting customers by purchasing behavior or grouping related documents before manual review.

Question 9

Why is data preprocessing important before training a model?

Accepted Answer

Data preprocessing is important because real datasets contain missing values, inconsistent formats, and features on very different scales, all of which can slow training or distort results. Cleaning and transforming the data helps the model focus on meaningful patterns rather than on artifacts of how the data was recorded.

Question 10

How should missing values be handled in a dataset?

Accepted Answer

Missing values can be handled by removing affected rows or columns, or by imputing a replacement such as the column mean, median, or most frequent category. The right choice depends on how much data is missing and why. Compute imputation values from training data only, so test information does not leak into training.

Question 11

What is feature scaling and why does it matter?

Accepted Answer

Feature scaling transforms numeric features to comparable ranges, for example by standardizing each feature to zero mean and unit variance or by rescaling values to a fixed interval. It matters because distance-based methods and gradient-based training can be dominated by features with large numeric ranges, such as income compared with age.

Question 12

What is one-hot encoding used for in machine learning?

Accepted Answer

One-hot encoding converts a categorical feature with several possible values into a set of binary columns, one per category, where exactly one column is set to one for each example. It lets models that expect numbers use categories such as color or country without implying a false numeric order among them.

Question 13

Why split data into training, validation, and test sets?

Accepted Answer

Splitting data lets you train on one portion, tune settings on a validation portion, and report final performance on a held-out test portion. Using separate data for each step keeps performance estimates honest, because a model judged on examples it helped shape will look better than it really is.

Question 14

What is data leakage in a machine learning project?

Accepted Answer

Data leakage occurs when information that would not be available at prediction time influences training, making results look better than they will be in practice. Examples include scaling using statistics from the full dataset or including a feature that is recorded only after the outcome occurs. Leakage is a common cause of disappointing deployments.

Question 15

What is feature engineering and why do practitioners perform it?

Accepted Answer

Feature engineering is the process of creating new input variables from raw data so that patterns become easier for a model to learn. Practitioners might combine columns, extract the day of the week from a timestamp, or take a logarithm of a skewed amount. Good features can improve results more than switching algorithms does.

Question 16

What does it mean to split data by time rather than at random?

Accepted Answer

Splitting by time means earlier records form the training set and later records form the test set, mirroring how a model will be used on future data. Random splits can let the model peek at the future, which inflates results for forecasting problems. Time-based splits give a realistic estimate whenever order matters.

Question 17

What is overfitting, and how can you recognize it?

Accepted Answer

Overfitting occurs when a model learns noise and quirks of the training data so closely that it performs poorly on new data. You can recognize it when training accuracy is high but validation or test accuracy is much lower. Regularization, more data, and simpler models are common remedies.

Question 18

What is underfitting and how does it differ from overfitting?

Accepted Answer

Underfitting occurs when a model is too simple or trained too little to capture the real pattern in the data, so it performs poorly even on training examples. Overfitting is the opposite problem, where the model fits training noise and fails to generalize. Both are checked by comparing training and validation performance.

Question 19

What is the bias-variance tradeoff in model selection?

Accepted Answer

The bias-variance tradeoff describes how total prediction error splits into bias, the error from overly rigid assumptions, and variance, the error from sensitivity to the particular training sample. Simple models tend toward high bias, while very flexible models tend toward high variance. Good models balance these two sources of error.

Question 20

Which metrics are commonly used to evaluate a classifier?

Accepted Answer

Common classification metrics include accuracy, precision, recall, and F1 score, which is the harmonic mean of precision and recall. Accuracy alone can mislead when classes are imbalanced, so precision and recall give a clearer view of errors. The confusion matrix is the table that supports all of them.

Question 21

How do precision and recall differ when judging a classifier?

Accepted Answer

Precision is the share of positive predictions that are actually correct, while recall is the share of actual positives that the model successfully finds. High precision means few false alarms, and high recall means few missed cases. The right balance depends on which kind of error costs more in the application.

Question 22

What is k-fold cross-validation and why would you use it?

Accepted Answer

K-fold cross-validation splits the training data into k parts, trains the model k times with a different part held out each time, and averages the scores. It gives a more stable estimate of performance than a single split, especially when data is limited. It is also a common way to compare candidate settings fairly.

Question 23

What is regularization and how does it help control overfitting?

Accepted Answer

Regularization adds a penalty to the loss function that discourages large or complicated parameter values. By favoring simpler solutions, it reduces the model's ability to memorize noise in the training data. Common forms include L1 and L2 penalties, which shrink weights in different ways.

Question 24

What does a learning curve show about a model's training?

Accepted Answer

A learning curve plots model performance against the amount of training data or training iterations, usually showing both training and validation scores. If both curves plateau at a low score, the model is likely underfitting. If training scores stay high while validation scores lag, more data or regularization may help.

Question 25

What is linear regression and what does it model?

Accepted Answer

Linear regression models a numeric output as a weighted sum of input features plus an intercept. Its weights show how each feature is associated with the target, holding other features fixed. It is simple, fast, and interpretable, which makes it a useful baseline for many prediction tasks.

Question 26

How does logistic regression differ from linear regression?

Accepted Answer

Logistic regression predicts the probability that an example belongs to a class by passing a weighted sum through a sigmoid function, which squashes values between zero and one. Despite its name, it is a classification method. Its output is usually converted into a class label by applying a threshold, such as 0.5.

Question 27

How does a decision tree make predictions for new examples?

Accepted Answer

A decision tree makes predictions by asking a sequence of questions about feature values, moving from a root node through branches until it reaches a leaf. Each leaf holds a predicted class or numeric value. Trees are easy to visualize and explain, but a single deep tree tends to overfit.

Question 28

What is k-nearest neighbors and how does it classify an example?

Accepted Answer

K-nearest neighbors classifies a new example by finding the k training examples closest to it under a distance measure, then taking a majority vote among their labels. It has no explicit training phase, so all the work happens at prediction time. Feature scaling matters greatly for this method.

Question 29

What is a support vector machine trying to find?

Accepted Answer

A support vector machine seeks the separating boundary, called a hyperplane, that leaves the widest possible margin between classes. The training examples closest to that boundary are the support vectors, and they alone determine its position. Kernel functions extend the method to nonlinear boundaries.

Question 30

How does the k-means clustering algorithm work in practice?

Accepted Answer

K-means chooses k cluster centers, assigns each example to its nearest center, recomputes each center as the mean of its assigned examples, and repeats until assignments stop changing. The number k must be chosen in advance. Results depend on the starting centers, so practitioners often run the method several times.

Question 31

What is a random forest and why does averaging help?

Accepted Answer

A random forest is an ensemble of many decision trees, each trained on a random sample of the data and considering a random subset of features at each split. Averaging their predictions, or taking a majority vote, reduces variance because individual tree errors partly cancel out. This usually yields better accuracy than one tree.

Question 32

What is gradient boosting and how does it build a model?

Accepted Answer

Gradient boosting builds an ensemble sequentially, adding one small model at a time. Each new model is trained to correct the errors, or more precisely the gradient of the loss, left by the models already combined. The final prediction is the weighted sum of all these models.

Question 33

Why is it important to check a model for bias against groups?

Accepted Answer

A model trained on historical data can reproduce or amplify unfair patterns present in that data, leading to worse outcomes for some groups of people. Checking performance separately across relevant groups helps reveal these gaps. Responsible practice also involves documenting the data source and deciding whether the model should be used at all.

Question 34

What is a baseline model and why should you build one first?

Accepted Answer

A baseline model is a simple reference, such as predicting the most common class or the average value, or a basic linear model. It shows how much a more complex approach actually improves results. Starting with a baseline also reveals whether the problem is solvable with the available data.

Question 35

What does it mean for a model to generalize?

Accepted Answer

Generalization means a model performs well on new examples drawn from the same underlying process, not just on the data it was trained on. It is the central goal of supervised learning. Evaluating on held-out data is the standard way to check whether it has been achieved.

Question 36

How should you monitor a model after it is deployed?

Accepted Answer

Monitoring means tracking prediction quality, input data distributions, and system performance over time. Real-world data often drifts away from the training data, which can quietly degrade accuracy. Teams typically set alerts and plan periodic retraining when metrics fall below an agreed threshold.

Question 37

What is data drift and why does it matter after deployment?

Accepted Answer

Data drift is a change over time in the statistical properties of the inputs a model receives, compared with the data used to train it. When drift occurs, relationships the model learned may no longer hold, so predictions become less reliable. Detecting drift early helps teams decide when to update the model.

Question 38

Why is it important to explain a model's predictions to stakeholders?

Accepted Answer

Explanations help people judge whether a prediction is trustworthy, spot errors, and decide whether acting on it is appropriate. Some methods, such as examining coefficients or feature importance, offer global views, while others explain individual predictions. Clear explanations also support accountability when decisions affect people.

Question 39

What is the difference between a parameter and a hyperparameter?

Accepted Answer

A parameter is a value the model learns from data during training, such as a linear regression weight. A hyperparameter is a setting chosen before training, such as the learning rate, tree depth, or the value of k. Hyperparameters are typically tuned by comparing performance on validation data.

Question 40

What is a good habit for keeping machine learning experiments reproducible?

Accepted Answer

Record the data version, code version, random seeds, hyperparameters, and evaluation results for every experiment. Fixing random seeds makes training runs repeatable, and version control keeps code changes traceable. Reproducible work lets others verify results and lets you reliably compare future changes.

Machine Learning Basics Flashcards

1Core Concepts of Learning

2Preparing and Understanding Data

3Training and Evaluating Models

4Common Algorithms

5Applied Practice and Responsible Use

Learning Resources