Multilevel Oversampler for Bias Reduction in Imbalanced Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models often become biased when trained with imbalanced data, leading to inaccurate predictions, particularly in critical applications like tumor detection, where false negatives can have severe consequences.
Innovation Solution
A method that involves generating oversampled training data by calculating distance values between minority and majority data points, using these values to determine a sample rate, and then oversampling the minority level data points to create a more balanced dataset, which is used to train the model, thereby reducing bias.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional oversampling techniques (e.g., SMOTE) are used to balance imbalanced data, then model accuracy improves, but computation time and energy consumption increase
Solution Approach 1:
The patent changes the parameter of oversampling strategy from uniform/random sampling to distance-based selective sampling. By calculating distance values between minority class samples and majority class samples, the system identifies and oversamples only those minority samples that are most critical (farther from majority class boundaries), thereby maintaining model accuracy while reducing the total number of oversampled samples and associated computation time
Solution Approach 2:
The patent applies local quality by treating different minority class samples differently based on their distance metrics. Instead of uniformly oversampling all minority samples, the system identifies specific regions in the feature space where minority samples are most vulnerable to misclassification and applies oversampling selectively to those local regions, optimizing both accuracy and efficiency
2Reliability
If traditional oversampling techniques (e.g., SMOTE) are used to balance imbalanced data, then model accuracy improves, but energy consumption increases
Solution Approach 1:
The patent modifies the oversampling parameter from comprehensive to selective based on distance calculations. By changing the sampling criterion to distance-based selection, the system reduces the computational burden and energy consumption associated with generating and processing oversampled samples, while still achieving the goal of improving model accuracy on imbalanced data
Solution Approach 2:
The patent applies partial action by performing oversampling only on the subset of minority class samples that fall below a distance threshold, rather than oversampling all minority samples. This partial approach suffices to improve model accuracy on critical regions while consuming less energy compared to exhaustive oversampling methods
3Loss of time
If distance-based selective oversampling is used, then computation time and energy consumption are reduced, but model accuracy may be compromised
Solution Approach 1:
The patent carefully tunes the distance threshold parameter to balance between oversampling intensity and computational efficiency. By adjusting this parameter, the system ensures that enough critical minority samples are oversampled to maintain model accuracy, while avoiding excessive oversampling that would increase computation time. The threshold is set to capture the most informative minority samples
Data Source
AI summary
Methods, systems, and apparatus, including computer programs encoded on computer-storage media, for oversampling training data for training a machine learning model. In some implementations, techniques described enable debiasing of trained models through oversampling of training data. In general, training data can be partitioned into multiple levels corresponding to specific elements of data, e.g., type of tumor, type of individual, time of day, among others. A given level can be associated with higher bias than other levels. For example, input data of a particular type of individual, such as an individual of a certain age range or with certain genetic traits, can cause corresponding machine learning model predictions or results that are more biased compared to other data. Higher bias can include an increased number of false positives or false negatives for Boolean model predictions or predictions.