Time-Based Ensemble Model for Anomaly Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional machine learning models for detecting abnormal behavior, such as money laundering, face limitations due to training data that may not capture short-term volatility, periodic retraining schedules that do not account for data changes, and reliance on a subset of features, leading to inaccurate predictions and increased false positives.
Innovation Solution
A time-based ensemble machine learning model that combines multiple machine learning models trained on different time periods, retrained based on feature volatility and prediction accuracy, to improve prediction accuracy by considering both short-term and long-term feature dynamics.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If machine learning models are trained on historical data, then prediction capability is established, but the models fail to capture short-term volatility and recent data changes
Solution Approach 1:
The patent implements dynamic retraining of machine learning models by continuously monitoring feature volatility and triggering retraining when thresholds are exceeded. The system dynamically adjusts the training data window based on data freshness and volatility, allowing models to adapt to changing patterns while maintaining the ability to capture short-term volatility and recent data changes.
Solution Approach 2:
The patent segments the training data into multiple time-based windows (e.g., recent 30 days, previous 60 days, older 90 days) and trains separate models or model components for each segment. This segmentation allows the system to handle different temporal patterns independently, improving both prediction accuracy for current patterns and adaptability to data changes over time.
2Reliability
If models are retrained periodically, then data freshness is maintained, but retraining occurs regardless of actual data changes causing resource waste
Solution Approach 1:
The patent implements a feedback mechanism that continuously monitors feature volatility metrics and triggers retraining only when volatility exceeds predetermined thresholds. This feedback-driven approach ensures retraining occurs only when data changes are significant enough to affect prediction accuracy, eliminating unnecessary computational resources while maintaining model reliability.
Solution Approach 2:
The patent performs preliminary analysis of data volatility before initiating retraining. By pre-calculating volatility metrics and establishing trigger thresholds in advance, the system avoids unnecessary retraining operations and optimizes computational resource allocation to only when and where needed.
3Device complexity
If a subset of features is used, then model complexity is reduced, but prediction accuracy decreases due to missed patterns
Solution Approach 1:
The patent applies local quality by differentiating the treatment of features based on their volatility and importance. High-volatility features receive enhanced weighting and dedicated analysis, while stable features are treated differently. This localized approach allows the model to focus computational attention on the most informative features for each prediction context, improving accuracy without uniformly increasing complexity across all features.
Data Source
AI summary
An input dataset is sorted into a first version of data and a second version of data. The first version of data is associated with a first period of time and the second version of data is associated with a second period of time. The second period of time is a shorter period of time than the first period of time. A first set of one or more machine learning models is generated based on the first version of data. A second set of one or more machine learning models is generated based on the second version of data. The first set of one or more machine learning models and the second set of one or more machine learning models are combined to generate an ensemble model. A prediction based on the ensemble model is outputted. The prediction indicates abnormal behavior associated with the input dataset.