Description
Date and Time:
June 25, 2024: 10:30 am – 11:30 am (JST)
Venue: Online and Open Space at the RIKEN AIP Nihonbashi office
Title:
Feature Interaction Selection (FIS) and its Applications in Structure Learning, Robust Generalization, and Interpretability
Speaker:
James Enouen
Abstract:
Generalized Additive Models (GAMs) have been the dominant approach for creating interpretable models across most data modalities. After briefly discussing the differences between interpretable approaches and explainable approaches, we highlight the key problem of Feature Interaction Selection (FIS) for GAM models and beyond. Its application across different models and modalities and the myriad benefits of this combinatorial learning problem are discussed in terms of robustness, interpretability, and learned structure. Insights for the future of interpretability and explainability are then discussed.
Bio:
James Enouen is a PhD student researching at University of Southern California under the advisement of Prof. Yan Liu on interpretability methods. Currently interning at NII with Prof. Mahito Sugiyama, he is working on tensor decomposition and information geometry. His main research area is in interpretable methods for machine learning and deep learning. His work is focused on marrying ideas from generalized additive models, high-dimensional statistics, causal inference, and structure learning.
