7 Introduction to Machine Learning

What machine learning does

Machine learning builds programs that learn patterns from data and use them to make predictions or organize information. Rather than writing a rule for every case, a developer supplies examples and a learning algorithm. A resulting model is useful when it performs well on examples it has not seen before.

Learning from labeled examples

In Supervised learning, each training example includes input information called Features and a desired answer called a Label or target. The model learns a relationship between the inputs and targets, then uses that relationship to predict targets for new examples.

A housing dataset might include home size, neighborhood, and number of bedrooms as features, with sale price as the target. A model could use those examples to predict prices for homes that have not yet sold.

Regression predicts a numeric value, such as a home's price or tomorrow's temperature. Mean absolute error (MAE) measures the average absolute difference between predicted and actual values.

Classification predicts a category, such as whether an email is spam or which plant species appears in a photo. A fraud model that predicts whether a transaction is fraudulent performs classification; predicting the transaction's likely dollar loss is regression.

Accuracy is the fraction of predictions that are correct. When classes are imbalanced, Precision, the fraction of positive predictions that are correct, and Recall, the fraction of actual positives found, can be more informative.

Finding structure without labels

In Unsupervised learning, examples have features but no supplied target labels. The algorithm looks for structure in the data. Its results can help people explore a dataset, but a discovered pattern is not automatically meaningful or useful.

Clustering groups examples according to similarity. For example, a retailer might group customers by purchase patterns to explore different shopping behaviors. Other unsupervised methods reduce the number of dimensions used to represent data, making patterns easier to inspect or use in later analysis.

Because there may be no known correct answer, unsupervised results often require interpretation. A cluster score can measure properties such as how compact or separated groups are, but the score alone does not prove that the groups have real-world importance.

Separating training and evaluation

A Training set is used to learn a model's patterns, while a held-out Test set estimates how well the trained model may perform on new examples. For example, a dataset might be divided into $80\%$ training data and $20\%$ test data; the exact split depends on the problem and the available data.

For supervised learning, keep each example's features together with its label when splitting the data. In classification, a Stratified split can help preserve the proportions of each class in both sets. A random split may give an unrealistic evaluation when observations are related, such as multiple records from the same person, or ordered in time. In those cases, keep related groups together or, when that better matches the model's use, train on an earlier period and test on a later one.

Preprocessing requires the same care. If missing values are filled in or features are scaled, learn those transformations using training data only, then apply them to the test data. Using test-set information during training or preprocessing causes Data leakage and can make evaluation results overly optimistic.

A Validation set or Cross-validation can help compare model choices during development. Keep the test set for a final evaluation: repeatedly adjusting a model based on its test score can gradually make the test set part of the development process.

Choosing and interpreting evaluation measures

Choose evaluation measures to match both the task and the consequences of errors. For regression, MAE gives average error in the target's units. Root mean squared error (RMSE) also measures prediction error, while giving relatively more weight to large errors.

For classification, accuracy gives the overall proportion of correct predictions. Precision and recall describe different aspects of positive predictions, so the appropriate balance depends on the application. A screening system may prioritize finding as many true cases as possible, while another application may need to limit false alarms.

For clustering, measures of within-group similarity and between-group separation can help compare results. Human review and domain knowledge are often needed to decide whether groups are useful. A score is only an estimate based on the chosen data, split, and metric; it does not guarantee that a model will perform equally well in every real-world setting.

Limits, overfitting, and bias

Overfitting happens when a model learns training examples too closely, including random noise or accidental details, and then performs poorly on new data. A much better score on training data than on held-out data is one warning sign. Simpler models, more representative training examples, or other appropriate controls can help, but the right approach depends on the task.

A model can also Underfit: it fails to capture useful patterns and performs poorly even on its training data. Evaluation can help reveal both problems, but scores should be interpreted alongside the data and task.

Bias can enter through the data, labels, measurement choices, or assumptions built into a system. If a training dataset underrepresents a population, contains historically biased decisions, or uses a misleading target, a model may repeat or amplify those problems. Overall performance can hide uneven errors across groups, so it can be important to examine data representation and evaluate results for relevant subgroups. No single metric or procedure guarantees fairness.

Supervised learning supports tasks such as regression and classification, while unsupervised learning looks for structure in unlabeled data. Training and test sets help distinguish learning from evaluation, and validation approaches support model selection. Careful metric choice and avoiding leakage are important, but strong test performance does not eliminate risks from overfitting, biased data, or changes in conditions.