OrdShap: Feature Position Importance for Sequential Black-Box Models

Fuente: arXiv
Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Hill, Davin, Hill, Brian L., Masoomi, Aria, Nori, Vijay S., Tillman, Robert E., Dy, Jennifer
Format: Preprint
Veröffentlicht: 2025
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866915599564668928
author Hill, Davin
Hill, Brian L.
Masoomi, Aria
Nori, Vijay S.
Tillman, Robert E.
Dy, Jennifer
author_facet Hill, Davin
Hill, Brian L.
Masoomi, Aria
Nori, Vijay S.
Tillman, Robert E.
Dy, Jennifer
contents Sequential deep learning models excel in domains with temporal or sequential dependencies, but their complexity necessitates post-hoc feature attribution methods for understanding their predictions. While existing techniques quantify feature importance, they inherently assume fixed feature ordering - conflating the effects of (1) feature values and (2) their positions within input sequences. To address this gap, we introduce OrdShap, a novel attribution method that disentangles these effects by quantifying how a model's predictions change in response to permuting feature position. We establish a game-theoretic connection between OrdShap and Sanchez-Bergantiños values, providing a theoretically grounded approach to position-sensitive attribution. Empirical results from health, natural language, and synthetic datasets highlight OrdShap's effectiveness in capturing feature value and feature position attributions, and provide deeper insight into model behavior.
format Preprint
id arxiv_https___arxiv_org_abs_2507_11855
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle OrdShap: Feature Position Importance for Sequential Black-Box Models
Hill, Davin
Hill, Brian L.
Masoomi, Aria
Nori, Vijay S.
Tillman, Robert E.
Dy, Jennifer
Machine Learning
Sequential deep learning models excel in domains with temporal or sequential dependencies, but their complexity necessitates post-hoc feature attribution methods for understanding their predictions. While existing techniques quantify feature importance, they inherently assume fixed feature ordering - conflating the effects of (1) feature values and (2) their positions within input sequences. To address this gap, we introduce OrdShap, a novel attribution method that disentangles these effects by quantifying how a model's predictions change in response to permuting feature position. We establish a game-theoretic connection between OrdShap and Sanchez-Bergantiños values, providing a theoretically grounded approach to position-sensitive attribution. Empirical results from health, natural language, and synthetic datasets highlight OrdShap's effectiveness in capturing feature value and feature position attributions, and provide deeper insight into model behavior.
title OrdShap: Feature Position Importance for Sequential Black-Box Models
topic Machine Learning
url https://arxiv.org/abs/2507.11855