João Fonseca

ExplainerPFN: Towards tabular foundation models for model-free zero-shot feature importance estimations

Thu, 01 Jan 2026 00:00:00 +0000

Computing the importance of features in supervised classification tasks is critical for model interpretability. Shapley values are a widely used approach f

Mixture-of-Expert Blocks Contain Strong Hallucination Detection Signals

Thu, 01 Jan 2026 00:00:00 +0000

Despite their widespread use, Large Language Models (LLMs) remain limited by a fundamental problem: the generation of plausible but false content, known as

RelShap: Relationally Consistent Shapley Explanations

Thu, 01 Jan 2026 00:00:00 +0000

Machine learning pipelines commonly flatten relational data into single-table representations, discarding structural constraints. Widely used Shapley value

Auditing and Safeguarding Large Language Models

Wed, 18 Jun 2025 00:00:00 +0000

This project focuses on developing comprehensive methods for auditing and safeguarding Large Language Models (LLMs) to ensure their safe and responsible deployment in real-world applications.

Key Components

SafeNudge

A real-time safeguarding method designed to protect Large Language Models against red teaming attacks and harmful prompt injections. SafeNudge provides tunable safety-performance trade-offs, allowing organizations to customize protection levels based on their specific use cases and risk tolerance.

This paper is currently under submission, an early preprint is available on ArXiv and here: SafeNudge: Real-Time Safeguarding for Large Language Models.

How Much Effort Is Enough? Fairness in Algorithmic Recourse Through the Lens of Substantive Equality of Opportunity

Wed, 01 Jan 2025 00:00:00 +0000

Algorithmic recourse, or enabling individuals to reverse a negative outcome, has gained attention as a means of supporting human agency in interactions wit

SAFENUDGE: Safeguarding Large Language Models in Real-time with Tunable Safety-Performance Trade-offs

Wed, 01 Jan 2025 00:00:00 +0000

Large Language Models (LLMs) have been shown to be susceptible to jailbreak attacks, or adversarial attacks used to illicit high risk behavior from a model

SHAP-based Explanations are Sensitive to Feature Representation

Wed, 01 Jan 2025 00:00:00 +0000

Local feature-based explanations are a key component of the XAI toolkit. These explanations compute feature importance values relative to an “interpretable

ShaRP: Explaining Rankings and Preferences with Shapley Values

Wed, 01 Jan 2025 00:00:00 +0000

Algorithmic decisions in critical domains such as hiring, college admissions, and lending are often based on rankings. Given the impact of these decisions

Machine Learning Explainability Frameworks

Thu, 03 Oct 2024 00:00:00 +0000

This project focuses on developing frameworks for explaining and interpreting machine learning model predictions.

Key Components

ShaRP (Shapley for Rankings and Preferences)

A framework that explains the contributions of features to different aspects of ranked outcomes, based on Shapley values. This approach is particularly useful for understanding how different features influence ranking decisions in recommendation systems, search results, and other ranked outputs.

See the paper for more details: ShaRP: Explaining Rankings with Shapley Values.

Location

Wed, 08 May 2024 00:00:00 +0000

João Fonseca's office location at the Faculdade de Ciências, Universidade de Lisboa, in Lisbon, Portugal.

Office Hours

Wed, 08 May 2024 00:00:00 +0000

Schedule and location for João Fonseca's office hours.

The Game Of Recourse: Simulating Algorithmic Recourse over Time to Improve Its Reliability and Fairness

Mon, 01 Jan 2024 00:00:00 +0000

Algorithmic recourse, or providing recommendations to individuals who receive an unfavorable outcome from an algorithmic system on how they can take action

Multi-agent Algorithmic Recourse Over Time

Sat, 01 Jul 2023 00:00:00 +0000

This project studies the importance of time in the reliability of algorithmic recourse. We highlight the lack of reliability in recourse recommendations over several competitive settings, potentially setting misguided expectations that could result in detrimental outcomes. These findings emphasize the importance of meticulous consideration when AI systems offer guidance in dynamic environments. Our paper, "Setting the Right Expectations: Algorithmic Recourse Over Time", won the Best AI Track Paper award at EAAMO'23!

Geometric SMOTE for imbalanced datasets with nominal and continuous features