4 Machine Learning Basics

How models learn from data

Machine learning builds software that learns patterns from data and uses them to make predictions or decisions. A model is the learned function, while an algorithm is the learning procedure that adjusts it. Training adjusts the model’s parameters to reduce a loss, a numerical measure of prediction error.

Data examples contain features, or inputs, such as the words in an email. In supervised learning, examples also have a label, the desired answer, such as “spam.” The goal is for a model to generalize: to perform well on new examples rather than simply memorize its training examples.

Prediction tasks and learning approaches

The kind of prediction determines the task. Regression predicts a numerical value: for example, estimating an apartment’s sale price from its size, location, and age. Classification predicts a category: for example, assigning an email to “spam” or “not spam.”

Learning approaches

• Supervised learning learns from labeled examples. It is used for classification and regression, such as predicting whether a transaction is fraudulent from past transactions labeled as fraudulent or legitimate.

• Unsupervised learning looks for structure without supplied answer labels. Clustering can group customers with similar behavior, while dimensionality reduction can represent complex data with fewer variables.

• Semi-supervised learning combines a smaller labeled set with a larger unlabeled set. It is useful when labeling data is costly.

• Reinforcement learning involves an agent taking actions in an environment and receiving rewards or penalties. The agent aims to improve its choices over time.

Splitting data and preventing overfitting

A dataset is commonly divided into separate subsets so that model development and evaluation use different examples. Each subset has a distinct role:

• The training set is used to fit the model’s parameters.

• The validation set is used during development to compare models or tune choices such as settings and features.

• The test set is held aside until the end to estimate performance on unseen data.

A model evaluated on its training examples can appear successful simply because it memorized them. This is overfitting: training performance is strong, but performance on new data is weaker. Repeatedly adjusting a model based on test results lets information from the test set influence development and makes the final estimate less trustworthy. Keep the test set separate until final evaluation.

The split must suit the data. For example, when predicting future demand, train on earlier dates and test on later dates rather than randomly mixing time periods. Avoid data leakage: information that would not genuinely be available at prediction time must not enter training or evaluation. Related or duplicate records should not appear across splits in a way that makes the test artificially easy. No single split ratio suits every dataset; each subset needs enough representative examples for its role.

When data is limited, cross-validation can help compare models. Divide the development data into several parts, train on all but one part, and validate on the held-out part. Repeat so each part serves as validation. Use these results for model selection, then evaluate the chosen model once on the untouched test set.

Evaluating model performance

Choose evaluation measures that match the task and the cost of different errors. Compare results with a simple baseline, such as always predicting the most frequent class or the average target value.

Classification measures

A confusion matrix counts correct and incorrect predictions. In binary classification, the counts are true positives, true negatives, false positives, and false negatives.

• Accuracy is the fraction of predictions that are correct. It can be misleading when one class is much more common than another.

• Precision is the fraction of predicted-positive examples that truly are positive. It is important when false alarms are costly.

• Recall is the fraction of truly positive examples that the model finds. It is important when missing positives is costly.

For example, when screening for a rare condition, a model that predicts “no condition” for everyone could have high accuracy but zero recall for people who have the condition. Accuracy alone would hide this failure.

Regression measures

For regression, mean absolute error (MAE) is the average absolute difference between predicted and actual values. Root mean squared error (RMSE) gives larger errors extra influence. The coefficient of determination, denoted by $R^2$, compares a model’s predictions with a simple mean-based baseline; interpret it alongside error measures and the task context.

A single score never captures every kind of model quality. Inspect relevant errors and use evaluation data that resembles the data the model will encounter.