Course contentsShow
← Machine Learning and Deep Learning
Lesson 373 of 3,5388. Unsupervised Learning: ClusteringFree lesson

What is Anomaly Detection?

Define anomalies, outliers, and novelties; understand when anomaly detection is needed vs. standard classification.

What is Anomaly Detection?

What you'll learn: How to identify unusual patterns in data that don't conform to expected behavior, and when anomaly detection is the right approach.

Understanding the Core Concept

Anomaly detection is the process of identifying data points, events, or observations that differ significantly from the majority of your data. Think of it like a quality control inspector on a production line: most products are normal, but occasionally something defective comes through that needs to be flagged.

In machine learning, we distinguish between three related concepts:

  • Anomalies (or outliers): Data points that deviate significantly from normal patterns (e.g., fraudulent credit card transactions)
  • Outliers: Extreme values that lie far from other observations (e.g., a billionaire in a salary dataset)
  • Novelties: New patterns not seen during training but not necessarily bad (e.g., a new type of network traffic)

When to Use Anomaly Detection vs. Classification

Use anomaly detection when:

  • Anomalies are extremely rare (highly imbalanced data)
  • You have mostly normal examples with few or no labeled anomalies
  • The types of anomalies are unknown or constantly evolving
  • Examples: fraud detection, equipment failure prediction, network intrusion

Use standard classification when:

  • You have sufficient labeled examples of both normal and abnormal classes
  • Anomaly patterns are well-defined and consistent
  • The problem is more balanced (not 99.9% normal cases)

Key Takeaway: Anomaly detection specializes in finding rare, unusual patterns when you have mostly normal data and limited or no labeled anomaly examples—situations where standard supervised classification struggles.