Anomaly Detection Model Segmentation for Time Series Precision
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine learning models for detecting anomalous events in time series data suffer from reduced accuracy and inconsistent results due to high false positive rates and the inability to effectively incorporate user feedback in a low latency runtime environment.
Innovation Solution
The proposed solution involves training a multitask model offline to predict anomaly event types and suitable model types for each event, selecting appropriate model instances with tuned hyperparameters, and updating these models in both offline and runtime environments based on user feedback.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If machine learning models are used to detect anomalous events in time series data, then detection capability is provided, but accuracy is reduced due to high false positive rates
Solution Approach 1:
The patent segments the anomaly detection task into multiple sub-tasks by detecting different anomaly types (point anomalies, contextual anomalies, collective anomalies) using specialized models for each type. This segmentation allows each model to focus on specific patterns, reducing false positives while maintaining comprehensive detection coverage.
Solution Approach 2:
The patent changes the parameters of the anomaly detection system by introducing multiple models with different hyperparameters and thresholds. Each model is tuned for specific anomaly types, allowing optimization of detection accuracy for each category while managing false positives through parameter-specific calibration.
2Reliability
If machine learning models are trained to detect anomalies, then detection functionality is achieved, but results are inconsistent
Solution Approach 1:
The system segments detection into multiple specialized models, each trained on specific anomaly types with consistent labeling guidelines. This segmentation ensures consistent detection behavior for each anomaly type while the overall system maintains comprehensive coverage through model specialization.
Solution Approach 2:
The patent implements feedback mechanisms where detection results are reviewed and labeled data is continuously updated. This feedback loop ensures consistent labeling practices and allows models to learn from previous detections, improving overall detection consistency and reliability.
3Measurement precision
If user feedback is incorporated to improve model accuracy, then detection precision improves, but system complexity increases
Solution Approach 1:
The patent segments the feedback incorporation process by maintaining separate labeled data stores for different anomaly types. This allows user feedback to be processed and applied to specific models without affecting the entire system, reducing overall complexity while maintaining precision improvements.
Solution Approach 2:
The system performs preliminary actions by pre-labeling data and pre-training models before incorporating user feedback. This preliminary preparation allows feedback to be integrated more efficiently, reducing the complexity of real-time updates while maintaining high detection precision.
Data Source
AI summary
Techniques for increasing the precision of machine learning models that are trained to detect anomalous data points in a time series. The techniques including methods and systems for training machine learning models offline, using the trained machine learning models to predict anomalies in an online runtime environment, and updating anomaly detection models in the runtime and offline environments. The machine learning models may include a multitask model for predicting one or more anomalous events present in input time series data, and for each identified anomalous event type, predicting a type of machine learning model that is best suited for predicting that anomalous event type. The models may further include model instances selected using the predicted one or more anomalous events present in the input time series data and the predicted anomalous event type and used to predict an anomaly event in the input time series data.