Training Data Downsampling for Fair Decision Boundary Preservation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional downsampling methods for training machine learning models do not adequately address data imbalance and often lead to fairness issues, resulting in inaccurate and unfair predictions, particularly when training datasets are unbalanced with respect to protected attributes.
Innovation Solution
A method involving categorizing data points into primary subsets based on attributes, selecting specific attribute values, and downsampling auxiliary subsets by retaining points with the smallest average distance to the furthest points in the primary subset, thereby generating a downsampled training dataset that maintains decision boundary integrity and improves fairness.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If conventional downsampling methods are used to address data imbalance, then the quantity of minority class samples is increased, but the accuracy and fairness of the ML model deteriorate due to loss of critical decision boundary information
Solution Approach 1:
The patent applies local quality by treating different regions of the feature space differently. Critical regions near decision boundaries are preserved with higher density, while non-critical regions are downsampled more aggressively. This is achieved by computing density ratios between majority and minority classes and selectively retaining samples based on their local density characteristics, ensuring that locally important samples are preserved while globally balancing the dataset.
Solution Approach 2:
The patent changes the parameter of sample selection from uniform random sampling to density-based selective sampling. By introducing density ratio computation and comparing it against threshold values, the method dynamically adjusts which samples are retained. This parameter change allows the downsampling process to adapt to local data characteristics, preserving samples that are critical for decision boundaries while removing redundant samples.
2Quantity of substance
If conventional downsampling methods are used to address data imbalance, then the quantity of minority class samples is increased, but the fairness of the ML model deteriorates due to biased representation of protected groups
Solution Approach 1:
The patent ensures fair representation by preserving local density characteristics within each protected group. By computing density ratios separately for majority and minority classes and retaining samples based on their relative density, the method maintains proportional representation of subgroups within protected attributes. This prevents the loss of locally important samples that could lead to biased predictions against protected groups.
3Quantity of substance
If downsampling is applied to balance training data, then data imbalance is reduced, but critical data points near decision boundaries are lost leading to underfitting
Solution Approach 1:
The patent prevents underfitting by preserving samples in high-density regions that are critical for defining decision boundaries. The density ratio computation identifies regions where samples are locally important, and these samples are retained even when overall downsampling is applied. This ensures that the model learns from critical boundary regions while still achieving balanced class representation.
Solution Approach 2:
The patent performs preliminary density analysis before downsampling to identify and protect critical samples. By computing density ratios and determining retention thresholds in advance, the method prepares a selection criterion that prioritizes samples near decision boundaries. This preliminary action ensures that critical information is preserved before the downsampling process removes redundant samples.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A method comprising: categorizing each datapoint in a training dataset into primary subsets based on first and second attributes; selecting a specific value of the first attribute and dividing each of the primary subsets corresponding to the selected value into a plurality of auxiliary subsets; for each of the primary subsets corresponding to the selected value, downsampling the plurality of auxiliary subsets with respect to the other primary subsets, respectively, to generate a plurality of downsampled auxiliary subsets, wherein the downsampling comprises: for each datapoint in the auxiliary subset concerned, computing an average distance to the k furthest datapoints of the primary subset concerned in respect of the plurality of attributes other than the at least first and second attributes; and removing n datapoints of the auxiliary subset concerned having the largest computed average distance to generate the downsampled auxiliary subset concerned, where k and n are positive integers.