Predictive machine learning

Begin by naming the output you need. A category calls for classification; a numerical target calls for regression. When the task has no labelled target, clustering groups observations by similarity, while dimensionality reduction represents data with fewer dimensions. These methods answer different questions, so a cluster should not be treated as a predicted class unless you have labels that support that interpretation.

A decision tree is a readable starting point for classification or regression: it applies rules based on features. A random forest combines predictions from randomized trees. Anomaly detection flags observations that differ from expected patterns; a recommender system selects or ranks items based on a user or context. These are candidate approaches, not guarantees of useful predictions. Define the target or decision first, then compare a simple baseline against methods that fit it.

• Decision Trees

  Explains how decision trees use feature-based rules for classification and regression, including interpretability and overfitting concerns.

• Ensembles: random forests

  Describes how randomized trees are combined in a forest.

• Collaborative filtering

  Introduces recommendations based on similarities among users and items.

Classification

Predicting which category an input belongs to.

Regression

Predicting a numerical value.

Clustering

Grouping items according to similarities in their data.

Dimensionality reduction

Representing data with fewer dimensions while retaining useful information.

Decision tree

A model that follows branching conditions to reach a prediction.

Random forest

A model combining predictions from multiple decision trees.

Anomaly detection

Identifying observations that differ from expected patterns.

Recommender system

A system selecting or ranking items for a user or situation.

Supervised learning uses examples with known target values or categories. For a classification task, begin by checking class frequencies and compare a model against a simple majority-class prediction. For a numerical target, compare a model against a simple constant prediction such as the training-set mean. These baselines make it easier to tell whether added complexity improves the task metric.

For a first model, linear or logistic regression gives a reference point; a small decision tree can provide a rule-based comparison. Keep model selection separate from the final test. Use training data to fit, validation data to choose, and test data once for a final estimate. A test set that is reused to tune choices can stop representing a clean check. Make the split resemble the way future cases will arrive, and check that held-out examples are representative and not duplicates of training examples.

• Datasets: Dividing the original dataset

  Explains training, validation and test sets and why repeated test-set use can distort evaluation.

• Linear regression

  Introduces a simple numerical prediction model and its relationship between features and a label.

When examples do not have labels, start by asking whether the aim is to find groups, compress a representation or flag unusual observations. Clustering can reveal patterns, but the result depends on how similarity is represented and should be checked against the question you are investigating. Anomaly detection also depends on what counts as expected: methods may learn from a typical set of observations or search for outliers already present in the data. Treat flags as signals to examine, not as explanations by themselves.

Forecasting differs because the target is in the future. A simple forecast can use recent values or a historical average as a reference, then be compared with a time-series method. Evaluate with rolling time-based splits: each training window should contain only observations that came before its test point. Choose a forecast horizon that matches the real decision, since one-step accuracy may not describe a multi-step use.

• Time series cross-validation

  Shows rolling-origin evaluation in which each training set uses only observations before its test point.

• Novelty and Outlier Detection

  Distinguishes outlier detection from novelty detection and explains their use for anomaly detection.