ML Ensemble Anomaly Detection for High-Dimensional KPI Monitoring
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional anomaly detection algorithms are ineffective in detecting small, meaningful anomalies in system metrics, known as 'slow bleed' anomalies, and struggle to display complex information in a way that allows users to effectively interpret and remediate issues in large datasets.
Innovation Solution
The system employs an ensemble of machine learning algorithms with a multi-agent voting system to detect anomalies and generates interactive visuals, such as radar-based and tree map visuals, to represent high-dimensional data sets, enabling users to identify problems in real-time and predict future states of the platform.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional anomaly detection algorithms are used, then large dips or spikes in metrics are detected, but small meaningful anomalies (slow bleed anomalies) are not detected
Solution Approach 1:
The system segments the anomaly detection task into multiple specialized algorithms, each optimized for different types of anomalies. The ensemble includes algorithms specifically designed to detect gradual changes and slow bleed anomalies, rather than relying on a single algorithm that only detects large spikes. This segmentation allows the system to simultaneously detect both obvious large anomalies and subtle slow bleed anomalies with high precision.
Solution Approach 2:
The system uses a composite approach by combining multiple machine learning algorithms into an ensemble. Each algorithm contributes its strengths to the overall detection system, creating a composite detection mechanism that is more robust and reliable than any single algorithm. The ensemble methodology integrates results from multiple algorithms to improve both precision and reliability for detecting various types of anomalies including slow bleed anomalies.
2Quantity of substance
If hundreds of metrics are monitored, then comprehensive system monitoring is achieved, but user ability to interpret and take action on the information is overwhelmed
Solution Approach 1:
The system merges information from hundreds of individual metrics into consolidated visual representations that highlight system-wide patterns and anomalies. Instead of presenting users with hundreds of separate metric displays, the system combines them into unified visualizations that show correlations and relationships across metrics, making the information more manageable and interpretable while maintaining comprehensive monitoring coverage.
Solution Approach 2:
The system transforms the high-dimensional data from hundreds of metrics into lower-dimensional visual representations that preserve essential information. By projecting the complex multi-metric data space into visual dimensions that humans can perceive and interpret, the system maintains the comprehensive monitoring capability while making the information accessible and actionable for users through intuitive visual displays.
3Adaptability or versatility
If independent user expertise is used for predicting system state, then human judgment is applied, but predictions are unreliable or inaccurate
Solution Approach 1:
The system implements feedback loops where machine learning algorithms continuously learn from historical data and anomaly patterns, improving their prediction accuracy over time. The system provides feedback to users about predicted system states and actual outcomes, allowing both the algorithms and users to refine their understanding. This feedback mechanism ensures that predictions become progressively more reliable while maintaining the adaptability of human judgment in interpreting and acting on predictions.
Data Source
AI summary
Described are computing systems and methods configured to detect a small, but meaningful, anomaly within one or more metrics associated with a platform. The system displays visuals of the metrics so that a user monitoring the platform can effectively notice a problem associated with the anomaly and take appropriate action to remediate the problem. An operational visual includes a radar-based visual with a heatmap arranging metrics, and a node representing a state of the metrics. Moreover, the system uses an ensemble of unsupervised machine learning algorithms for multi-dimensional clustering of hundreds of thousands of monitored metrics. Via the visuals and the implementation of the machine learning algorithms, the described techniques provide an improved way of representing and simulating many metrics being monitored for a platform. Moreover, the techniques are configured to expose actionable and useful information associated with the platform in a manner that can be effectively interpreted.