Event

Proposal Exam: Faisal Hamman

Tuesday, November 12, 2024
6:00 p.m.
Virtual
Maria Hoo
301 405 3681
mch@umd.edu

ANNOUNCEMENT: Ph.D. Research Proposal Exam

Name: Faisal Hamman

Committee:

Professor Sanghamitra Dutta (Chair)

Professor Sennur Ulukus

Professor Furong Huang

Date/Time: Monday, November 12, 2024 at 6:00 PM

Location: https://umd.zoom.us/j/3986469938

Title: Trustworthy and Explainable Machine Learning Using Information Theoretic Methods

Abstract: 

The growing use of machine learning (ML) in critical domains such as finance, healthcare, and education generates an urgent need for explainability and trust. This proposal seeks to address emergent challenges in explainable machine learning by developing novel mathematical frameworks deep-rooted in information-theoretic methods. 

An emerging problem in explainability is the problem of robust counterfactual explanations. Counterfactual explanations are changes to input features that would alter the prediction of a machine learning model, providing actionable insights into how an individual could achieve a desired outcome (algorithmic recourse). However, machine learning models are frequently updated, leading to a phenomenon of model multiplicity (also called Rashomon Effect) where equally well-performing models make conflicting predictions on the same inputs. Model multiplicity can cause potential invalidation of previously provided counterfactual explanations. How do we provide reliable algorithmic recourse while also being robust to model multiplicity (Rashomon Effect)?  We propose a stability measure to compute the resilience of counterfactuals to model changes along with probabilistic guarantees on validity. We also develop optimization algorithms to ensure that counterfactuals remain valid over time with experiments on benchmark tabular datasets.  

Building on our ideas, we next focus on the problem of consistency in predictions in TabLLMs (large language models fine-tuned for tabular data tasks). We first demonstrate the issue of fine-tuning multiplicity in TabLLMs where models fine-tuned with varying seeds can lead to inconsistent predictions despite having similar performance. To tackle this issue, we propose a consistency measure that mathematically quantifies the stability of predictions without requiring retraining multiple models, with both theoretical guarantees and empirical validation.

Lastly, shifting from model multiplicity to data heterogeneity, we address the emergent challenge of explaining fairness in federated learning environments, where data heterogeneity across clients creates discrepancies in not just model performance but also in fairness across protected groups. We study how global fairness (fairness across all clients) and local fairness (fairness specific to each client) interact under these conditions, using an information-theoretic tool called Partial Information Decomposition.

Paper Links:

1. Hamman,  S. Dutta,  "Demystifying Local and Global Fairness Trade-offs in Federated Learning Using Partial Information Decomposition," International Conference on Learning Representations (ICLR 2024).
2. Hamman,  S. Dutta,  "A Unified View of Group Fairness Tradeoffs Using Partial Information Decomposition,"  IEEE International Symposium on Information Theory (ISIT 2024). 
3. Hamman, E. Noorani, S. Mishra, D. Magazzeni, and S. Dutta, "Robust Algorithmic Recourse Under Model Multiplicity with Probabilistic Guarantees,” Journal on Selected Areas in Information Theory: Information-Theoretic Methods for Trustworthy Machine Learning (JSAIT 2024). 
4. Hamman, E. Noorani, S. Mishra, D. Magazzeni, S. Dutta, "Robust Counterfactual Explanations for Neural Networks With Probabilistic Guarantees," In: International Conference on Machine Learning (ICML 2023). 
5. Hamman, J. Chen, S. Dutta, “Can Querying for Bias Leak Protected Attributes? Achieving Privacy With Smooth Sensitivity,” In: ACM Conference on Fairness, Accountability, and Transparency (FAccT 2023).
6. Hamman, P. Dissanayake, S. Mishra, F. Lecue, S. Dutta, "Quantifying Prediction Consistency Under Model Multiplicity in Tabular LLMs”,arxiv.org/abs/2407.04173.

Audience: Faculty