fairness
Explain the denial
It is 2018, and Europe's new data-protection regime has put a question mark over every deployed model: a small-business owner denied a loan by a gradient-boosted ensemble demands to know why, and the bank's data scientists can produce only feature-importance folklore. Deliver explanation with substance: local attributions faithful to the model rather than merely plausible, counterfactual statements — what minimal change would flip the decision — that are actionable and honest about causal status, and stability, because an explanation that changes when the input wiggles is testimony no court should hear. Then confront the trade-off directly: when does the mandate justify an interpretable model outright, at measured accuracy cost? Get it wrong and "the algorithm decided" becomes an unappealable verdict — or explanations become reassurance theater laundering arbitrary decisions.
Who this problem belongs to
The two figures whose methods fit it best, out of 30 in contention.
Guestrin co-invented LIME, Local Interpretable Model-agnostic Explanations (2016, with Ribeiro and Singh), the exact technique this scenario needs: a method for producing local attributions faithful to what a black-box model, including a gradient-boosted ensemble, actually did on a specific input rather than plausible-sounding folklore about feature importance. His research directly engages the stability problem the scenario flags, since an explanation method whose output changes when the input wiggles slightly is testimony no court should trust, and LIME's own follow-up literature grapples with exactly this fragility. His work on XGBoost's distributed training also gives him firsthand authority over the model class under scrutiny. No other carrier on this list built the actual explanation methodology this problem centers on.
Pearl's causal framework, from Bayesian networks through do-calculus and his later work on counterfactual reasoning, is precisely the missing rigor this scenario's second demand requires: a counterfactual statement, what minimal change would flip the loan decision, is meaningless without being honest about its causal status, and Pearl is the researcher who formalized exactly when a counterfactual claim is warranted by the model versus merely a correlational artifact dressed up as actionable advice. His insistence on distinguishing seeing from doing gives the solver the exact conceptual tool to avoid the trap of offering recourse that sounds actionable but has no causal grounding. He did not build the explanation-interface tools themselves, keeping him just below Guestrin.
In the mind map
The same ideas, as concepts rather than history — in John's ML knowledge map.
30 figures are scored on this problem. Draw it in a battle to see where you land.