Skip to main content
Back

UCD's ISF Algorithm Finally Reveals Why AI Approves Your Loan

A new XAI method upgrades bland 'even if' explanations by surfacing the hidden features that quietly hold a decision in place.

Pro

Paper, Full Access
  • New ISF algorithm generates semi-factual XAI explanations that also reveal the hidden features holding a decision in place.
  • Existing 'even if' methods find extreme feature changes that preserve outcomes but never explain why the outcome survives.
  • ISF uses a seesaw pattern: key-feature contribution weakens while a hidden feature's contribution strengthens simultaneously.
  • Tested on 38,223 queries across Adult, Blood Alcohol, Diabetes, German Credit, and HELOC, beating ensemble baselines.
  • ISF hit near-100% seesaw coverage on 4 of 5 datasets while keeping distance, sparsity, and plausibility competitive.
  • User studies confirmed people prefer elaborated explanations; paper available on arXiv.

When a loan app tells a customer "even if you had asked for double, you would still have been approved," it sounds reassuring but leaves the real question unanswered: why? A new algorithm from University College Dublin, called Informative Semi-Factuals (ISF), tries to close that gap in explainable AI by exposing the hidden features that keep a decision stable when other inputs are stretched to extremes.

Semi-factuals have become a popular strategy in XAI, explaining how a prediction can hold steady even when certain inputs are altered. They mirror counterfactuals, which spell out the smallest change that would flip the decision. Where counterfactuals use "If only..." reasoning, semi-factuals use "Even if..." reasoning, telling users which feature-differences leave the outcome unchanged rather than which ones reverse it.

The missing 'why' in even-if explanations

Authors Saugat Aryal and Mark Keane point out a blind spot in a field they helped define with their earlier IJCAI 2023 survey. Most semi-factual methods compute explanations built around maximal value changes, typically to a single key feature of a query instance, that leave the outcome unaltered. Their earlier Most Distant Neighbor (MDN) method, for example, scans each feature dimension of the query for the furthest same-class instance in the dataset. None of these approaches address why it is possible to radically change a key feature's value without causing a class change.