NNG-Mix: Improving Semi-Supervised Anomaly Detection With Pseudo-Anomaly Generation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Anomaly detection (AD) is crucial for identifying rare events in complex systems.
• AD is often unsupervised due to high labeling costs, but semi-supervised approaches benefit from limited expert-labeled data.
• Existing augmentation techniques may not optimally leverage limited labeled anomalies.

Purpose of the Study:

• To introduce a novel algorithm, Nearest Neighbor Gaussian Mix-up (NNG-Mix), for generating pseudo-anomalies.
• To enhance semi-supervised and supervised anomaly detection by augmenting limited labeled anomaly data.
• To improve the performance of anomaly detection systems through effective data augmentation.

Main Methods:

• Developed the Nearest Neighbor Gaussian Mix-up (NNG-Mix) algorithm.
• NNG-Mix integrates labeled and unlabeled data to generate pseudo-anomalies.
• Evaluated NNG-Mix by training existing AD algorithms on augmented datasets and comparing with Mixup and Cutout.

Main Results:

• NNG-Mix significantly outperforms common augmentation techniques like Mixup and Cutout.
• Experiments on 57 ADBench datasets show substantial performance improvements over baseline methods.
• Achieved up to 16.4% improvement on Classical, 8.8% on CV, and 8.0% on NLP datasets in ADBench.

Conclusions:

• NNG-Mix is an effective data augmentation method for semi-supervised and supervised anomaly detection.
• The algorithm successfully generates valuable pseudo-anomalies, enhancing AD model performance.
• NNG-Mix offers a practical solution for improving anomaly detection with limited labeled data.