Supervised machine learning learns a mapping from inputs to known outputs. Inputs are features, outputs are labels, and the adjustable values inside the model are parameters.

During training, the model makes predictions on examples whose labels are known. A loss function measures how wrong those predictions are. The training algorithm changes the parameters to reduce that loss.

Linear regression is a model family. Gradient descent is one method that can be used to find its parameters.

A model is not the training algorithm

Understand features, labels, parameters, training, and inference.

What a model actually learns

Build the mental model behind training, evaluation, loss functions, and predictions before implementing ML algorithms.

Performance on training data tells you how well a model fit examples it already saw. What matters is generalization: performance on new examples drawn from the same real-world process.

Leakage happens when information from validation or test data influences training. Calculating normalization statistics on the complete dataset is a common example—the model indirectly learns about held-out examples.

Measure whether a model generalizes instead of memorizing its examples.

Train, validation, and test sets

A loss function turns model quality into a number that can be minimized. For regression, mean squared error penalizes the squared distance between predictions and targets.

The gradient describes how the loss changes when each parameter changes. Gradient descent takes a small step in the opposite direction of that gradient.

Training is repeated feedback: predict, measure error, compute a direction, update parameters, and repeat.

The big picture

See how a numeric error signal becomes a better set of model parameters.

Loss and gradient descent

A binary classifier produces a score or probability that is converted into a class using a threshold. Every prediction becomes a true positive, true negative, false positive, or false negative.

Accuracy can be misleading for imbalanced datasets. A fraud model that predicts “not fraud” for every transaction may be highly accurate while finding no fraud at all.

Use recall when missing a positive is especially costly. Use precision when false alarms are especially costly.

Start with the cost of errors

Choose accuracy, precision, recall, or F1 based on the cost of mistakes.

Classification and metrics

Libraries such as scikit-learn provide reliable implementations of standard algorithms. Your responsibility is to define the problem, prepare data correctly, avoid leakage, and interpret the evaluation.

You now have the vocabulary and mental model needed to implement individual algorithms and metrics in the JoinAI problem catalog.

Next step

Put preprocessing, training, prediction, and evaluation into one clean workflow.