Random Forest

A random forest is a machine learning algorithm that works by creating a large number of decision trees — each trained on a slightly different random subset of the data — and then combining their predictions. The "forest" is the collection of trees, and the final answer is determined by majority vote (for classification) or averaging (for regression).

How a single decision tree works

A decision tree is like a flowchart. It asks a series of yes/no questions about your data to arrive at a prediction. For example, a tree predicting customer churn might ask: Is the customer's contract month-to-month? Have they called support more than three times? Is their monthly bill above £80? Each answer leads to the next question until the tree reaches a prediction.

The problem with a single decision tree is that it tends to memorise the training data too closely — a problem called overfitting. It performs brilliantly on data it has seen but poorly on new data.

How the forest fixes this

A random forest solves overfitting by building hundreds or thousands of trees, each slightly different:

• Each tree is trained on a random sample of the training data (with replacement)
• At each decision point, each tree considers only a random subset of available features
• The final prediction is the consensus of all trees

This randomness makes the forest robust. Individual trees might make mistakes, but their errors tend to cancel out when aggregated.

When random forests are used

Random forests excel at structured data problems — spreadsheets, databases, and tabular data. Common applications include:

• Credit scoring and fraud detection
• Customer churn prediction
• Medical diagnosis support
• Feature importance analysis (identifying which variables matter most)

Strengths and limitations

Random forests are fast to train, hard to break, and require minimal tuning. They handle missing data well and naturally rank feature importance. However, they struggle with unstructured data like images and text, where deep learning methods are superior. They also produce models that are harder to interpret than a single decision tree.