Anomaly Detection

(Master of Information Science)

• It is different from tasks like Class classification/Regression because it detects things that have no previous examples.
  • In terms of Predictive and Discovery-based Machine Learning, it is considered a discovery-based approach.
• It searches for outliers.
• It is also effective to use clustering to group past data and consider anything that does not fit into any group as an anomaly.

(Master of Information Science) Machine Learning

• Anomaly Detection
  • k th-NN
    • First, measure the distance between your point and all other points.
    • Sort them and use the distance to the k-th nearest point as the score.
    • Why k-th?
      • We want to find outliers.
      • If k=1, we may not be able to find outliers that are close to each other.
      • So we use k=2, 3, or higher.