Ilie Sarpe "How do classifiers interact? Exact attribution for plurality-voting ensembles"

This talk can be viewed via zoom. (Note: this talk will be recorded)

ABSTRACT
Modern practical and accountable data analysis algorithms face several challenging issues, such as scalability on massive data and the rigorous theoretical guarantees a user requires to trust the output of an algorithm.
I will briefly introduce the key principles guiding my research on the foundations of algorithmic data analysis, and the areas on which I am focusing.
Then, I will present in detail my approach through the recent NeurIPS 2026 paper 'Exact power indices for plurality-voting ensembles', in the area of trustworthy ML.
The work studies the attribution problem for ensemble models that aggregate multiple base classifiers through plurality voting, with ties broken uniformly at random, as in popular software libraries (e.g., scikit-learn or R).
We use cooperative game theory to define a new game that captures the expected accuracy of an ensemble, and we formulate the attribution problem as the computation of the Shapley and Banzhaf power indices.
We then design ENPOWER, a suite of polynomial-time algorithms to compute the exact power indices under our new game. 
Our methods exploit the combinatorial formulation of the new cooperative game together with the use of ordinary generating functions to obtain the power indices for multiclass classification.
The resulting algorithms are extremely efficient (i.e., often requiring less than 2 seconds on practical inputs), and can be used for applications such as prompt ensembling of vision models.
I will conclude by discussing future directions of my research in the area.

BIO
Ilie Sarpe is a researcher in the foundations of algorithmic data analysis. He received his PhD from the University of Padova in 2023, advised by Prof. Fabio Vandin, and recently completed a postdoc at KTH Royal Institute of Technology, mentored by Prof. Aristides Gionis.
His research focuses on the design of practical, scalable algorithms with rigorous provable guarantees, in areas such as graph mining, optimization and, more recently, explainable machine learning.
His work has appeared at venues that include KDD, WWW, VLDB, CIKM, NeurIPS, and WSDM.