Unified Anomaly Score Thresholding for Contamination Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing unsupervised anomaly detection (UAD) algorithms struggle to accurately determine the contamination factor in datasets without labeled data, as they often misinterpret anomaly scores and lack the ability to adapt to diverse datasets and models, leading to ineffective anomaly detection.
Innovation Solution
A meta-learning approach that combines anomaly score unification with a meta-learning method to estimate contamination factors, using a universal anomaly threshold calculated as an average of anomaly thresholds from multiple labeled datasets, allowing adaptation to new datasets and models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If unsupervised anomaly detection algorithms are used to identify anomalies without labeled data, then the ability to detect anomalies is improved, but the accuracy in determining contamination factor deteriorates
Solution Approach 1:
The patent introduces anomaly score unification as an intermediary step that transforms raw anomaly scores from different detectors into a standardized form. This unified score distribution serves as a mediator that enables accurate contamination factor estimation without requiring labeled data, thus resolving the contradiction between unsupervised detection capability and measurement precision.
Solution Approach 2:
The patent transforms the parameter space by converting diverse anomaly scores from multiple detectors into a unified score distribution with standardized parameters. This parameter transformation allows the system to maintain adaptability across different detectors while achieving precise contamination factor determination through the learned relationship between unified scores and contamination levels.
2Ease of operation
If contamination factor is specified by users based on dataset percentage, then the anomaly detection process is simplified, but the reliability of detection deteriorates because users do not know the actual anomaly percentage
Solution Approach 1:
The patent implements a feedback mechanism where the system learns the relationship between anomaly score distributions and actual contamination factors from labeled datasets. This learned knowledge is then fed back into the detection process, allowing the system to automatically determine accurate contamination factors without relying on user guesses, thus maintaining ease of operation while improving reliability.
3Device complexity
If anomaly score distributions are assumed to be similar across datasets, then the complexity of handling diverse datasets is reduced, but the measurement precision of contamination factor deteriorates due to distribution differences
Solution Approach 1:
The patent segments the problem into two distinct parts: first, transforming diverse anomaly scores into a unified standardized distribution; second, learning the relationship between unified score distributions and contamination factors. This segmentation allows the system to handle dataset diversity systematically without sacrificing precision, as each segment addresses a specific aspect of the challenge.
4Reliability
If multiple anomaly detectors are used to improve detection quality, then the reliability of anomaly detection is improved, but the complexity of calibrating and unifying their scores deteriorates
Solution Approach 1:
The patent creates a universal anomaly score unification framework that can handle scores from multiple different anomaly detectors through a single standardized transformation process. This universal approach eliminates the need for detector-specific calibration procedures, allowing the system to leverage multiple detectors for improved reliability while maintaining simple, consistent processing complexity.
Data Source
AI summary
Herein is a universal anomaly threshold based on several labeled datasets and transformation of anomaly scores from one or more anomaly detectors. In an embodiment, a computer meta-learns from each anomaly detection algorithm and each labeled dataset as follows. A respective anomaly detector based on the anomaly detection algorithm is trained based on the dataset. The anomaly detector infers respective anomaly scores for tuples in the dataset. The following are ensured in the anomaly scores from the anomaly detector: i) regularity that an anomaly score of zero cannot indicate an anomaly and ii) normality that an inclusive range of zero to one contains the anomaly scores from the anomaly detector. A respective anomaly threshold is calculated for the anomaly scores from the anomaly detector. After all meta-learning, a universal anomaly threshold is calculated as an average of the anomaly thresholds. An anomaly is detected based on the universal anomaly threshold.