Distance-Based Downsampling for Fair ML Decision Boundaries
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional downsampling methods for training machine learning models fail to optimize dataset balance and often deteriorate fairness, leading to inaccurate and unfair predictions, particularly in imbalanced datasets with 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, ensuring that the downsampled dataset maintains critical decision boundary information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If conventional downsampling methods are used to address imbalance in training data, then the balance of the training dataset is improved, but the fairness and accuracy of the ML model deteriorate
Solution Approach 1:
The patent applies local quality by treating different regions of the dataset differently. Specifically, it identifies and preserves data points near decision boundaries (critical regions) while allowing more aggressive downsampling in other regions. This is achieved by calculating distances to furthest points and selectively retaining points with smaller average distances, thereby maintaining local fidelity in important areas while achieving global balance.
Solution Approach 2:
The patent changes the parameter of data point selection from random or uniform sampling to distance-based sampling. By computing the average distance from each point to its k furthest points in the same primary subset, and using this distance metric to guide which points to retain, the method transforms the downsampling process from a uniform parameter approach to a variable parameter approach that adapts to the local data structure.
2Quantity of substance
If data points are removed from the training dataset to achieve balance, then the dataset size is reduced, but important decision boundary information is lost
Solution Approach 1:
The patent identifies that not all data points are equally important. By calculating the average distance from each point to its k furthest points within the same primary subset (defined by protected attribute values), it identifies points with small average distances as being near decision boundaries. These critical points are preserved while other points are removed, thus maintaining decision boundary information while reducing dataset size.
Solution Approach 2:
The patent performs preliminary analysis of the data structure by categorizing points into primary subsets based on protected attributes and computing distance metrics before the actual downsampling occurs. This preliminary characterization of which points are important (near boundaries) versus which can be removed allows the method to make informed decisions about which points to retain, preventing loss of critical decision boundary information.
Data Source
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.