Synthetic Minority Data Augmentation for Imbalanced Event Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning methods struggle to predict low-incidence events, such as opioid overdoses, due to imbalanced data sets, where oversampling and undersampling techniques lead to overfitting and data loss, respectively, and synthetic data augmentation methods like SMOTE are ineffective in high-dimensional spaces.
Innovation Solution
A computer-implemented method generates synthetic minority-class training records by analyzing original minority-class records and creating new records by changing feature values based on predicted probabilities, thereby augmenting the training data without overfitting or losing information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If oversampling is used to balance classes, then the model can learn from more minority class examples, but the model tends to overfit and learn to identify instances of only those copied records
Solution Approach 1:
The patent creates synthetic copies of minority class records through data augmentation, generating new training examples that preserve the characteristics of the original minority class while providing variation. This allows the model to learn from more diverse examples without simply memorizing copied records.
Solution Approach 2:
The patent modifies parameters of existing minority class records to generate synthetic examples, changing feature values while maintaining the underlying class characteristics. This creates diverse training examples that improve generalization while preserving the minority class signal.
2Quantity of substance
If undersampling is used to balance classes, then the classes become more balanced, but data is thrown away which hurts the model when it tries to classify unseen data
Solution Approach 1:
Instead of discarding majority class data, the patent creates synthetic copies of minority class examples to achieve class balance. This preserves all original data while adding augmented minority class examples, avoiding information loss.
Solution Approach 2:
The patent transforms existing minority class records into new synthetic examples by modifying parameters, thereby increasing minority class representation without removing any majority class data.
3Quantity of substance
If SMOTE is used for data augmentation, then synthetic minority class records are generated, but the method does not work well in high-dimensional space because the concept of nearness breaks down
Solution Approach 1:
The patent extracts and modifies individual features of minority class records independently, rather than relying on high-dimensional proximity calculations. This avoids the curse of dimensionality by operating on individual feature dimensions rather than treating the entire high-dimensional space as a unified metric space.
Solution Approach 2:
The patent segments the data augmentation process into individual feature modifications, applying transformations to each feature separately. This breakdown of the high-dimensional problem into lower-dimensional feature-level operations avoids the limitations of SMOTE in high-dimensional spaces.
Data Source
AI summary
A method for generating synthetic training records for use in training a model to predict low-incidence events. A synthetic training record is generated from a minority-class training record by substituting a different value for a feature in the minority-class training record, where the probability of the different value occurring in the minority-class training record exceeds a probability threshold. Also disclosed are a non-transitory storage medium comprising minority-class training records and synthetic training records and a method of training a machine-leaning model using training records augmented with synthetic training records. An exemplary synthetic training records is a synthetic medical record for use in training a model to predict drug overdoses.