1 Introduction to Machine Learning and Datasets

How machine learning works

Machine learning (ML) builds computer programs that learn patterns from examples and use those patterns to predict outcomes or organize new data. Rather than specifying every rule by hand, a developer provides data and chooses a learning method. The resulting model captures patterns that can be applied to new examples, and its usefulness depends in part on whether its data represents the problem it will face.

For example, instead of writing rules for every kind of spam email, examples of emails and their known categories can be provided. A model can learn associations between email characteristics and categories, then predict a category for a new email.

Common machine learning tasks

The kind of output a model should produce helps determine the machine learning problem type. Regression predicts a numerical quantity, such as a home’s sale price or tomorrow’s temperature. Classification predicts a category, such as whether an email is spam or not spam; categories can be binary or have several possible classes.

Clustering groups examples by similarity when groups or labels are not provided in advance. For instance, a store might group customers with similar purchasing patterns.

Regression and classification typically use labeled examples, making them supervised learning tasks. Clustering is a common unsupervised task: it looks for structure in data without target labels.

Dataset structure and quality

A dataset is a collection of examples. In a common tabular format, each row represents one example and each column records a property of that example. Columns used as model inputs are called features; in supervised learning, the value to be predicted is the label or target.

In a home dataset, each row might describe one home. Size, bedrooms, and neighborhood are features, while sale price is the target for a regression task. If the task instead predicts whether a home sold above $350,000, the target is a category and the task is classification.

Features can be numerical, such as size, or categorical, such as neighborhood. Datasets can also contain text, images, audio, and other forms of data. Missing values, duplicate examples, incorrect entries, or wrong labels can reduce data quality, so datasets should be checked before use.

Training and testing data

For supervised learning, a training set contains examples with features and known targets that the model uses to learn patterns. A separate test set is held back to assess how well the trained model predicts on examples it did not train on.

For the assessment to be meaningful, test examples should represent the intended use and should not duplicate training examples. Keeping training and test data separate helps assess performance on unseen examples.