Atsushi Koike proposes simpler pairwise interactions for interpretable tabular prediction

Atsushi Koike proposes simpler pairwise interactions for interpretable tabular prediction

Atsushi Koike argues that a model can be visually transparent yet still too complicated for practitioners to interpret or check against domain knowledge. Generalized Additive Models with pairwise interactions (GA²Ms) display each feature pair as a contour plot, but Koike says methods such as EBM and NODE-GAM can learn shape functions with irregular contours and sharp local fluctuations. BAM, or Bivariate Additive Models, constrains each pairwise shape function to an asymmetric bivariate Gaussian, giving each component a unimodal contour map. Training selects the feature pairs automatically. Koike’s aim is stronger auditability: interactions simple enough to inspect, while retaining accuracy if the right feature pairs are captured. In experiments on six tabular benchmarks, he reports that BAM’s predictive accuracy is comparable to standard GAMs and approaches that of GA²Ms. The paper also shows a trade-off between accuracy and the number of components to inspect. The abstract reports benchmark results but does not establish how well BAM works beyond them or whether practitioners find its visualizations meaningfully easier to audit.

Publication: 8317ed6d-392c-46d3-8cf8-48b07a28e181

attribution — pass: The paper is correctly attributed to its sole listed author, Atsushi Koike, and arXiv is the venue. The argument and results are kept as Koike's claims. Minor note: the draft uses 'he', which the source does not state. Neutral phrasing such as 'the author reports' would avoid that assumption.

Authors:Atsushi Koike

summary — uncertain: The draft is accurate but may not be readable to non-specialists. Central terms are left unexplained in plain language: 'tabular' (spreadsheet-style data), 'shape function' (the learned curve or surface showing how features affect a prediction), and 'unimodal' (a single peak). These terms carry the core idea. The draft also softens the source's force. The source says practitioners 'cannot meaningfully interpret' these plots, but the draft says methods 'can learn' complex functions and claim 1 says 'can be difficult'. Smallest repair: briefly gloss these terms and restore 'cannot meaningfully interpret'.

so complex (irregular contours and sharp local fluctuations) that a practitioner cannot meaningfully interpret them or check them against domain knowledge

date — pass: The originalPublished date, 2026-11-04, matches the v1 submission date. It does not use the later feed date of 2026-11-05.

[Submitted on 4 Nov 2026]

support — pass: Each claim matches the abstract. The accuracy claim is correctly scoped to the six benchmarks, as comparable to standard GAMs and approaching GA2Ms. The conditional 'if the right feature pairs are captured' is preserved. The limitations correctly say the abstract offers no user study and no named benchmarks. The question follows from the evidence and is not leading.

We argue that such simplification can match the accuracy of standard GAMs if the right feature pairs are captured.

media — pass: The selected media is the verified direct link to the arXiv abstract page, which is the original source. No better source media was found.