Recourse Path Generation with Differential Privacy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for generating recourse paths in machine learning models face challenges in ensuring privacy, realism, and actionability, particularly when providing multi-step paths for achieving favorable outcomes.
Innovation Solution
A platform, language, cloud, and database agnostic recourse path generating module is developed, which implements an end-to-end differentially private pipeline. This pipeline uses Differentially Private clustering to represent subsets of the private dataset, forming a graph with cluster centers as nodes to generate private and realistic recourse paths.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If counterfactual explanations are provided for single-step recourse, then privacy risks increase, but the simplicity and directness of the explanation is maintained
Solution Approach 1:
The patent segments the recourse path into multiple intermediate steps between the original instance and the counterfactual explanation. Instead of providing a direct single-step recourse that exposes private data, the system generates a sequence of gradual transformations through intermediate instances, each step being less sensitive than the full counterfactual. This segmentation maintains privacy guarantees while still providing actionable recourse guidance.
Solution Approach 2:
The patent introduces intermediate instances as mediators between the original private data and the final counterfactual explanation. These intermediate instances serve as privacy-preserving proxies that gradually transform the original instance toward the desired outcome without directly exposing sensitive information. The intermediary steps act as buffers that maintain privacy while enabling recourse generation.
2Reliability
If noise is added to recourse paths to ensure privacy, then privacy is protected, but the realism and actionability of the path deteriorates
Solution Approach 1:
The patent employs dynamic noise addition strategies where the amount and type of noise applied to intermediate instances varies based on their position in the recourse path and their sensitivity level. Rather than uniformly adding noise throughout, the system dynamically adjusts privacy perturbations to maintain realism in less sensitive regions while ensuring privacy in more sensitive transitions. This dynamic approach preserves actionability while maintaining privacy guarantees.
Solution Approach 2:
The patent applies differential privacy measures locally to specific intermediate instances based on their sensitivity and position in the recourse path. Rather than uniformly applying noise to all points, the system selectively applies privacy protections where needed while leaving other regions with higher realism requirements less perturbed. This localized approach maintains overall path realism while ensuring privacy where critical.
3Adaptability or versatility
If multi-step recourse paths are generated, then the comprehensiveness and actionability improve, but the computational complexity and processing time increases
Solution Approach 1:
The patent performs preliminary actions by pre-computing and caching intermediate instances and their transformations before generating final recourse paths. The system prepares a library of privacy-preserving intermediate states in advance, which can be efficiently retrieved and combined during recourse generation. This preliminary preparation reduces the computational burden during actual recourse generation while maintaining comprehensive multi-step path quality.
Solution Approach 2:
The patent segments the complex multi-step recourse generation process into manageable sub-tasks that can be independently computed and optimized. By dividing the overall recourse path into smaller intermediate transformation steps, the system can process each segment separately using efficient algorithms, then combine them into complete recourse paths. This segmentation reduces memory requirements and enables parallel processing while maintaining comprehensive coverage.
Data Source
AI summary
Various methods and processes, apparatuses/systems, and media for generating realistic multi-step recourse paths while preserving privacy of customers are disclosed. A processor trains an ML model by using the at least a first set training data; implements a data distribution sampling algorithm on the first set of training data to generate corresponding sampled data by partitioning the first set of training data into non-overlapping subsets with differentially private clustering; computes a plurality of cluster centers with differential privacy guarantees for each of said plurality of cluster centers; generates a graph that connects each cluster center with different weights between each cluster center; and automatically generates, for a data point that receives a negative outcome from the trained model, a recourse path with privacy guarantees based on a plurality of set points from the graph that provides shortest path to output a positive outcome.