Naive Bayes Classification

Description

A fast, simple classifier that predicts categories (such as ticket topic or churn risk) from features assumed independent.

Naive Bayes estimates, for each class, the probability of each feature value, then multiplies them to score a new case. The "naive" assumption that features are independent is usually wrong, but the ranking of classes is often still good. It trains in seconds, needs little data and is easy to explain. In UX research it is useful for auto-tagging open-ended survey responses or support tickets into known codes after a human has coded a training set, and as a quick baseline before more complex models.

Attributes

Loss type
Unmigrated
Stage
Post-launch
Data
Mixed
Effort
Hours

Steps

  1. Code a training set by hand (for example 500 verbatims into themes).
  2. Convert text or categorical predictors into features (words, answers).
  3. Estimate class priors and conditional probabilities, with smoothing for unseen values.
  4. Predict classes for new cases and evaluate with a confusion matrix on held-out data.
  5. Review misclassified examples and refine codes or features.

Sources

  • Practical Statistics for Data Scientists (Bruce & Bruce)

Also known as

  • Naive Bayes

Categories

  • Analytics & Behaviour
  • Computational & Machine Learning