Anomaly Detection Quiz

Answers and explanations appear as you take the quiz.

1. Why are anomaly detection models often trained mostly on 'normal' data rather than labeled anomalies?
   • Normal data is illegal to label
   • Models cannot process anomalies at all
   • Normal data requires no computation
   • Anomalies are rare and labeled examples are scarce
2. How does an Isolation Forest identify an anomaly?
   • Anomalies are isolated in fewer random splits, giving shorter path lengths
   • Anomalies require many random splits to isolate
   • It measures the color of each data point
   • It clusters points by their labels first
3. When an autoencoder is used for anomaly detection, what signals an anomaly?
   • Low reconstruction error
   • High reconstruction error on unfamiliar input
   • A faster training time
   • More layers in the network
4. What does the Local Outlier Factor method primarily rely on?
   • The alphabetical order of features
   • The total file size of the dataset
   • The density of points around a given point
   • A fixed list of known fraud cases
5. Which is a realistic challenge for deployed anomaly detection systems?
   • Anomalies never change over time
   • Normal data is always perfectly labeled
   • Outliers are always the majority of data
   • The definition of 'normal' can drift, requiring adaptation
6. Which application is a classic use of anomaly detection?
   • Detecting fraudulent credit card transactions
   • Sorting photos alphabetically
   • Translating text between languages
   • Compressing video files
7. What does the contamination setting commonly represent in an Isolation Forest workflow?
   • The number of training trees that contain audio
   • An assumed or expected proportion of outliers used for thresholding
   • The probability a record was encrypted
   • The average depth of a normal point
8. Why should anomaly scores be reviewed with domain context?
   • An anomaly score always proves fraud
   • All normal data has an identical distribution
   • Rare but valid events can look unusual, and errors have different costs
   • A model score replaces an investigation