Machine Learning Outlier Removal for Model Accuracy Improvement
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deployed machine learning models often experience a decline in accuracy over time, making it challenging to improve model accuracy through re-training with new data.
Innovation Solution
A method for identifying and removing preliminary outliers from transactions based on a derived cut-off value, allowing for the identification of training outliers that impact specific metrics, and re-labeling these outliers to generate re-training data that meets an absolute goal for improving model performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If re-training the model with new training data is performed to improve model accuracy, then model accuracy can be improved, but the process becomes challenging and complex
Solution Approach 1:
The patent extracts and removes outlier transactions from the training dataset before re-training the model. By identifying and eliminating data points that deviate significantly from normal patterns (outliers), the training process becomes simpler and more effective, as the model focuses on learning from representative data rather than being confused by anomalous cases.
Solution Approach 2:
The patent performs preliminary data cleaning by removing outliers before the actual model re-training process. This preliminary action of filtering the training data ensures that the subsequent re-training operates on high-quality data, reducing the complexity and improving the effectiveness of the overall process.
2Measurement precision
If outliers are removed from transactions to improve model accuracy, then model performance improves, but the process requires identifying and deleting preliminary outliers based on derived cut-off values
Solution Approach 1:
The patent replaces manual or complex statistical outlier detection methods with a machine learning-based approach. An initial model is trained to automatically identify and flag outlier transactions, substituting complex manual analysis with an automated computational system that can efficiently detect anomalies based on learned patterns.
Solution Approach 2:
The system performs self-service by automatically identifying and removing its own outliers through the trained model. The model uses its learned understanding of normal transaction patterns to autonomously detect deviations, eliminating the need for external manual intervention in the outlier identification process.
3Adaptability or versatility
If multiple metric values are determined for transactions to identify training outliers, then the ability to improve specific metrics is enhanced, but the process becomes more complex
Solution Approach 1:
The patent segments the training process by separating outlier identification and removal as a distinct preliminary step before model re-training. This segmentation allows the complex task of improving multiple metrics to be broken down into manageable phases: first clean the data by removing outliers, then proceed with targeted re-training for specific metric improvements.
Solution Approach 2:
The patent performs preliminary outlier removal based on derived cut-off values before conducting metric-specific model improvements. This preliminary action simplifies subsequent metric optimization by ensuring the training data is clean and representative, reducing the complexity of achieving targeted metric improvements.
Data Source
AI summary
Embodiments are disclosed for a method for machine learning model training outliers. The method includes determining multiple metric values for corresponding transactions generated by a machine learning model. The method also includes deleting multiple preliminary outliers from the transactions based on a derived cut-off value. Further, the method includes identifying an absolute goal for improving a metric of the machine learning model. Additionally, the method includes identifying multiple training outliers from the remaining transactions. The remaining transactions include the transactions remaining after deleting the preliminary outliers. Also, a metric value of the remaining transactions meets the absolute goal.