Passer au contenu

/ Department of Computer Science and Operations Research

Je donne

Rechercher

Navigation secondaire

Farzaneh Heidari's Predoc III Presentation

Dear all / Bonjour à tous,

We are happy to invite you to Farzaneh Heidari's Predoc III defense on Tuesday, September 5th, at 10am. (Hybrid event.)

Vous êtes cordialement invité.e.s à la présentation du sujet de recherche de Farzaneh Heidari, le mardi 5 septembre à 10h00. (Présentation hybride).


Title: Towards Explaining Graph Models Using Tensor Networks

Date: September, 5th, 2023 at 10:00am-12:00pm EST

Location: Auditorium 1 - 6650 rue Saint Urbain

 

Jury

PrésidentSridhar, Dhanya
Directeur de rechercheRabusseau, Guillaume
Co-Directeur de rechercheTang, Jian
Membre
Wolf, Guy

 

Abstract

While global model interpretability remains challenging, focusing on local explanations - understanding individual predictions - can be more tractable and equally valuable. Our interpretable local surrogate (ILS) method approximates the behavior of a black-box graph model by fitting a simple surrogate model in the local neighborhood of a given input example. Leveraging the interpretability of the surrogate, ILS is able to identify the most relevant nodes contributing to a specific prediction. To efficiently identify these nodes, first we utilize group sparse linear models as local surrogates. Through empirical evaluations on explainability benchmarks, our method consistently outperforms state-of-the-art graph explainability methods. This demonstrates the effectiveness of our approach in providing enhanced interpretability for GNN predictions. In ongoing work, we use adaptive tensor networks as an interpretable local surrogate. Tensor networks, originally conceptualized within the realm of quantum physics, offer a compact and structured representation for complex multidimensional data. By harnessing their capability to efficiently represent and factorize high-dimensional functions, we propose a mechanism to approximate the behavior of graph models in the vicinity of a particular instance.

Our methodology involves the creation of a tensor network that locally approximates the graph model around a specific instance. The tensor network acts as a surrogate function, offering a local and interpretable representation of the model's behavior, by representing higher-order interactions explicitly.

In the future work, by evaluating separation rank and entanglement entropy between regions of input, we hope to find the functional groups in novel molecules and introduce a quantum-centric avenue in molecule design.