Anomaly Detection and Optimization GMLE Practice Questions (Page 7)

Questions 31–35

1. 31application · medium

   A deep learning model for time-series forecasting shows high accuracy on training data but poor accuracy on validation data. The team suspects overfitting. Which combination of techniques is most appropriate to reduce overfitting while maintaining model capacity?

   • AApply dropout to hidden layers and use early stopping on validation loss.
   • BRemove all regularization and train for more epochs.
   • CUse only L1 regularization and increase batch size.
   • DIncrease the number of layers and use a higher learning rate.
   Reveal Solution Discussion Select an answer first
2. 32expert · hard

   A financial institution is deploying an anomaly detection model to flag suspicious transactions. The model is an isolation forest trained on historical transaction data. The institution has a regulatory requirement to catch at least 95% of fraudulent transactions, but the fraud investigators can only review 500 alerts per day. The model currently achieves 90% recall with 500 alerts per day. Which action should the team take?

   • ALower the anomaly score threshold to increase recall, accepting that the number of alerts will increase.
   • BRetrain the model with a higher contamination rate, which will make it more sensitive to anomalies.
   • CUse a different algorithm, such as a one-class SVM, which may have better recall.
   • DRaise the anomaly score threshold to reduce false positives, which will increase precision but may lower recall.
   Reveal Solution Discussion Select an answer first
4. 33expert · hard

   A team is deploying an anomaly detection system for industrial equipment. The cost of a missed failure is extremely high, but the maintenance team can only investigate a limited number of alerts per day. The model's ROC-AUC is 0.95, but at the current threshold, precision is 40% and recall is 90%. The team wants to reduce the number of false positives without significantly reducing the number of true positives caught. Which action is most appropriate?

   • AUse precision-at-k to select the top k alerts that maximize the number of true positives within the review capacity.
   • BRetrain the model with a different random seed to improve the ROC-AUC.
   • CLower the decision threshold to increase recall further.
   • DRaise the decision threshold to improve precision, accepting a lower recall.
   Reveal Solution Discussion Select an answer first
5. 34foundation · easy

   In a fraud detection system where the model outputs a risk score for each transaction, which metric is most useful when the investigation team can only review the top 100 flagged transactions per day?

   • AF1-score
   • BPrecision-at-k
   • CRecall
   • DROC-AUC
   Reveal Solution Discussion Select an answer first
6. 35expert · hard

   A machine learning team is optimizing an autoencoder for anomaly detection on high-dimensional sensor data. They have a limited budget of 100 training runs. The hyperparameters include the number of hidden layers, the size of the bottleneck layer, the learning rate, and the regularization strength. The team wants to find the best configuration for detecting rare anomalies. Which optimization strategy is most appropriate?

   • AGrid search over a small set of values for each hyperparameter, ensuring all combinations are explored.
   • BRandom search over the full ranges of all hyperparameters, running 100 random combinations.
   • CManual tuning, starting with the team's best guess and adjusting based on validation performance.
   • DBayesian optimization, using a surrogate model to guide the search and focus on promising regions.
   Reveal Solution Discussion Select an answer first