Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Additive Models Explained: A Computational Complexity Approach
Authors: Shahaf Bassan, Michal Moshkovitz, Guy Katz
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Theoretical | Our work uncovers a nuanced spectrum of results, demonstrating that the complexity of obtaining explanations for GAMs is significantly influenced by (i) the input space domain (enumerable discrete, discrete, and continuous), (ii) the type of explanation (sufficient, contrastive, shapley value explanations, etc.), (iii) the underlying component model (e.g., neural networks, boosted trees, splines, etc.), (iv) and the distinction between classification and regression settings. Our findings reveal a broad range of unexpected and substantial complexity variations, driven by surprising factors such as the input domain which are not apparent in other ML models. We believe that our work lays the foundation for a wide range of implementations by enabling the development of tractable algorithms for computing explanations for GAMs in diverse settings. At the same time, it advances the theoretical understanding of these models by identifying when such explanations are computationally feasible and when they are not. |
| Researcher Affiliation | Collaboration | Shahaf Bassan The Hebrew University of Jerusalem EMAIL Michal Moshkovitz Google Research EMAIL Guy Katz The Hebrew University of Jerusalem EMAIL |
| Pseudocode | Yes | Algorithm 1 Cardinally Minimal Contrastive Reason Search Input f, x 1: Compute all values pi for every i [n] 2: F {1, . . . , n} Features for iteration 3: S The current sufficient reason 4: Sort F in descending order by the value of vi := βi xi pi for each i 5: for each i F do 6: if sign(P i S βi fi(xi) + P j S pj + β0 sign(P i n βi fi(xi) + β0) < 0 then 7: Break 8: end if 9: S S {i} 10: end for 11: return S S is a cardinally minimal contrastive reason |
| Open Source Code | No | This is a theoretical work; therefore, providing open access to data and code is not relevant. |
| Open Datasets | No | This is a theoretical work; therefore, providing open access to data and code is not relevant. |
| Dataset Splits | No | This is a theoretical work; thus, experimental reproducibility is not applicable. |
| Hardware Specification | No | This is a theoretical work; thus, information about computational resources is not applicable. |
| Software Dependencies | No | The paper does not mention any specific software names with version numbers used for its own work. It only provides examples of such software within the context of related work or general knowledge. |
| Experiment Setup | No | This is a theoretical work; therefore, hyperparameter details are not relevant. |