SHAP explanations can differ across repeated runs of the same estimator, even for a fixed model, input, and prediction. We characterize explanation multiplicity with a dual-seed protocol that compares model-induced and explainer-induced variability, magnitude-, rank-, and set-based metrics, and randomized reference baselines. Across datasets, models, and sampling strategies, multiplicity persists even for high-confidence predictions, while small magnitude differences can mask substantial changes in top-ranked features.
@inproceedings{hwang2026multiplicity,title={Explanation Multiplicity in SHAP: Characterization and Assessment},author={Hwang, Hyunseung and Lee, Seungeun and Rosenblatt, Lucas and Whang, Steven Euijong and Stoyanovich, Julia},booktitle={Advances in Neural Information Processing Systems (NeurIPS), Track on Evaluations and Datasets},year={2026},}
2025
FAccT
SHAP-based Explanations are Sensitive to Feature Representation
Hyunseung Hwang, Andrew Bell, Joao Fonseca, and 3 more authors
In Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency (FAccT), 2025
Local feature-based explanations such as SHAP are central to validating, documenting, and contesting high-stakes models, including the credit and lending classifiers used throughout financial services. We show that routine data-engineering choices, such as how a continuous feature like age is bucketed or how race is encoded, can substantially shift SHAP feature-importance rankings for the same model, and that this can be exploited to make a discriminatory model appear neutral. The findings bear directly on model risk management, fair-lending and adverse-action compliance, and whether a feature-importance explanation can be trusted at all.
@inproceedings{hwang2025shap,title={SHAP-based Explanations are Sensitive to Feature Representation},author={Hwang, Hyunseung and Bell, Andrew and Fonseca, Joao and Pliatsika, Venetia and Stoyanovich, Julia and Whang, Steven Euijong},booktitle={Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency (FAccT)},pages={1588--1601},year={2025},doi={10.1145/3715275.3732105},}
2023
AAAI
XClusters: Explainability-first Clustering
Hyunseung Hwang and Steven Euijong Whang
In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2023
Most explainable-clustering methods explain clusters only after the fact. XClusters instead makes explainability a first-class objective of the clustering itself, jointly minimizing cluster distortion and the size of the decision tree that explains the clusters, which matters wherever segments must be both accurate and defensible, as in customer and credit-risk segmentation (we evaluate on a real credit-card dataset). We show the objective is a difference of monotonic functions and give an efficient branch-and-bound algorithm that trades off accuracy against explanation simplicity.
@inproceedings{hwang2023xclusters,title={XClusters: Explainability-first Clustering},author={Hwang, Hyunseung and Whang, Steven Euijong},booktitle={Proceedings of the AAAI Conference on Artificial Intelligence (AAAI)},year={2023},}
2021
DMAH
Open-World COVID-19 Data Visualization
Hyunseung Hwang and Steven Euijong Whang
In VLDB Workshop on Data Management and Analytics for Medicine and Healthcare (DMAH), 2021
An extended abstract, from KAIST’s broader COVID-19 response, on the practical difficulty of producing useful visualizations in an open-world setting where it is not even clear up front which datasets exist or are relevant, a recurring obstacle in any data-driven decision pipeline.
@inproceedings{hwang2021covid,title={Open-World COVID-19 Data Visualization},author={Hwang, Hyunseung and Whang, Steven Euijong},booktitle={VLDB Workshop on Data Management and Analytics for Medicine and Healthcare (DMAH)},pages={81--84},year={2021},doi={10.1007/978-3-030-71055-2_8},}