arXiv cs.AI· Atsushi Koike·· 12 小时前AI 评分18
BAM：面向可解释表格预测的二元加性模型
BAM: Bivariate Additive Models for Interpretable Tabular Prediction
AI 导读
研究者提出二元加性模型（BAM），将 GA²M 中每个成对形状函数约束为非对称二元高斯这一简单参数形式，使每个分量呈现为单峰等高线图，特征对在训练中自动选择。在六个表格基准上，BAM 的预测精度与标准 GAM 相当，并接近 GA²M 的水平。研究还展示了精度与需检查分量数量之间的权衡。
正文 · 原文
Abstract:Generalized Additive Models with pairwise interactions (GA$^2$Ms) are regarded as glass-box models, since each pairwise shape function can be visualized as a contour plot. In practice, the shape functions learned by methods such as EBM or NODE-GAM are so complex (irregular contours and sharp local fluctuations) that a practitioner cannot meaningfully interpret them or check them against domain knowledge. Auditability in this stronger sense requires substantially simpler shape functions. We argue that such simplification can match the accuracy of standard GAMs if the right feature pairs are captured. We propose Bivariate Additive Models (BAMs), a class of GA$^2$Ms in which every pairwise shape function is constrained to a simple parametric form, instantiated as an asymmetric bivariate Gaussian. Each component is visualizable as a unimodal contour map, and its feature pair is selected automatically during training. Experiments on six tabular benchmarks show that BAM attains predictive accuracy comparable to standard GAMs and approaches that of GA$^2$Ms. We also show the trade-off between accuracy and the number of components to inspect.
| Subjects: | Machine Learning (cs.LG); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2611.05940 [cs.LG] |
| (or arXiv:2611.05940v1 [cs.LG] for this version) | |
| https://doi.org/10.48550/arXiv.2611.05940 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Atsushi Koike Dr. [view email]
[v1]
Wed, 4 Nov 2026 04:28:59 UTC (810 KB)
来源：arXiv cs.AI · arxiv.org