Interpretable models used in medicine and finance are usually additive. They predict with a sum of one-feature curves, so each feature's effect can be plotted and audited on its own. The cost is that such models cannot represent any effect that depends on two features jointly. A drop in nasal pressure may be harmless when blood oxygen is high and dangerous when it is low, and an additive model simply cannot say this (Vetter et al.). Tree ensembles have the opposite problem. Every root-to-leaf path is a conjunction of splits, so interactions come for free. But trees must stack splits on several features even to approximate a purely additive function, so counting which features co-occur on paths reports interactions that do not exist (Interaction Detection with Bayesian Decision Tree Ensembles).
We propose to let a random forest do the search over feature sets and to calibrate what it finds against a null forest. The null forest is grown on synthetic data that share the real inputs and the fitted additive structure but contain, by construction, no further interactions. A feature set enters the interpretable model only if the real forest uses it together far more often than forests grown under this "additive is enough" hypothesis. We repeat the procedure stage by stage, refitting the model and re-testing at the next interaction order. The result is a sparse additive model with a small set of 2-way and 3-way tree terms. It has no neural teacher and no MCMC, and it gives an explicit rate for admitting spurious terms.
Core Claim
The co-occurrence statistics of an unconstrained forest carry real interaction signal, but they are biased by an amount that depends on the data: feature correlations, main-effect shapes, and tree hyperparameters. That bias can be estimated directly. We refit the current additive model, generate responses equal to its predictions plus wild-bootstrapped residuals (residuals with random sign flips, which keeps any dependence of noise size on the inputs), and regrow the same forest on the unchanged inputs. The resulting distribution of co-occurrence scores is what hallucination looks like on this particular dataset.
Concretely, let $\mathrm{DWP}(S)$ be the depth-weighted prevalence of feature set $S$. This is the frequency with which the features in $S$ appear together along forest decision paths, weighted by depth, as defined in Provable Boolean interaction recovery from tree ensemble obtained via random forests. Let $\mathrm{DWP}^{(b)}_0(S)$ be the same quantity in the $b$-th of $B_0$ null forests. We admit $S$ when
$$\mathrm{DWP}(S) > q_{1-\alpha/M}\big(\{\mathrm{DWP}^{(b)}_0(S)\}_{b=1}^{B_0}\big),$$
where $q$ is an empirical quantile, $\alpha$ is the target rate of admitting any spurious term, and $M$ is the number of candidate sets the real forest proposed.
The novelty is not the use of forest co-occurrence, which LSSFind already analyzes. It is replacing LSSFind's asymptotic, effect-size-free threshold $2^{-|S|}$ with a finite-sample null tailored to the dataset, and then embedding the admitted sets as tree terms in an additive booster.
Relation to Prior Work
Several papers motivate the gap. Accurate intelligible models with pairwise interactions showed that adding a few pairwise terms to boosted-tree additive models cuts RMSE by 34% and classification error by 44% on average over plain additive models. Its FAST screen is limited to pairs, however, and places a single cut per feature. Sparse Interaction Additive Networks showed that 3-way terms help too, but finds them with a pretrained neural teacher and the Archipelago detector under a heredity rule. Heredity means a triple is tested only if all its pairs already interact, and anything the teacher misses is lost. Rosasco et al. show that features with weak main effects but strong joint effects are invisible to additive selection. This is exactly the case where a heredity-free proposer matters.
Other papers constrain the design. The Bayesian tree work shows that raw co-occurrence cannot be trusted: standard BART reports 1.5 to 2 spurious interactions on the Vo-Pati no-interaction benchmark. DP-Forests fixes this with a Dirichlet-process prior that groups trees into specialists, at MCMC cost and with sensitivity to the concentration parameter. ANOVA-BART achieves identifiable functional-ANOVA components with constrained trees and a stepwise heredity move, also via MCMC. LSSFind proves that co-occurrence recovers Boolean conjunctions, but it needs exponentially more samples as interaction order grows, is fragile to the per-split feature-sampling parameter $m_{try}$, and stops at mining rather than building a predictor.
We borrow three techniques. We take the DWP statistic from LSSFind and the cyclic boosting of shallow trees restricted to feature subsets from GA²M-style models. We reuse purification (moving main-effect signal out of interaction surfaces) as done in NODE-GAM, so that admitted terms are reported in a standard form.
One paper cuts against us. Beyond Additive Decompositions argues that some data are better described as products than as sums of low-order terms. On socmob its product model reports RMSE 9.48, against XGBoost at 17.65 and the Explainable Boosting Machine (EBM) at 20.21. Our model class stays additive in low-order terms, so we treat multiplicative data as a regime where we expect to lose on the raw scale, and we test explicitly whether a log link closes the gap.
Research Program
The procedure starts from a main-effects-only EBM-style model. At each stage it does four things. First, it fits a random forest to the current residuals and to the raw target, running both variants as an ablation. Second, it extracts the top $M$ candidate sets of order two and three by DWP. Third, it builds $B_0$ null forests (initially 50) on the current model's predictions plus sign-flipped residuals, with inputs, hyperparameters and random-seed schedule held identical to the real forest. Fourth, it admits the sets that pass the test above, adds a boosted shallow-tree term for each, refits cyclically, purifies, and repeats until nothing is admitted.
The forest proposes sets directly from its paths, so the procedure never enumerates all $\binom{p}{3}$ triples and never requires heredity. Its reach is still bounded by what greedy trees can find. Pure parity-style interactions with no marginal signal are out of scope, and we will say so.
We will analyze three questions. The first is when the wild-bootstrap null is valid: under an additive truth with noise independent of the inputs given the fit, it controls spurious admissions up to the error in estimating the additive part. The second is how misfit in that additive part leaks into false interactions, which we will study with deliberately under-smoothed main effects. The third is how the admission rate depends on $m_{try}$. We expect the calibration to absorb $m_{try}$ sensitivity, because the null forests share the same $m_{try}$.
The compute cost is roughly $B_0$ forest fits per stage. For Higgs-scale data we will grow null forests on subsamples and report how admitted sets change with subsample size.
Baselines
The naive reference is a main-effects-only EBM, the additive model the procedure starts from. Any admitted term must improve held-out loss over it. It also defines zero false interactions.
The strongest existing method for accuracy with pairwise interpretability is GA²M with FAST screening (Lou et al.), run through its EBM implementation. Its reported reference is the 34% average RMSE reduction over additive models. We will give it the same number of admitted terms our method selects and the same tuning budget.
The strongest competitor for higher-order interpretable accuracy is SIAN. Its reported average rank across its seven-dataset suite is 3.00, against 3.71 for NODE-GAM. We will run it with its own neural teacher and with the same maximum order of three.
The strongest competitor for false-interaction control is DP-Forests, reported to keep near-zero false positives on Vo-Pati no-interaction where BART finds 1.5 to 2. We will give it the same inputs and a wall-clock budget matched to our $B_0$ forest fits.
The unconstrained accuracy ceiling is XGBoost, tuned with the same hyperparameter search budget as our booster. For California Housing the best reported interpretable figure in the file is RMSE 0.451 ± 0.002 from Neural Additive Experts. No reported result was found on Lacuna for any of these methods on Bike Sharing or MIMIC in numerical form.
Three ablations isolate the key ingredient. The first uses the same forest and DWP ranking but admits sets by LSSFind's fixed $2^{-|S|}$ bound. The second uses the same ranking but admits the top-$k$ sets, with $k$ chosen by validation loss. The third keeps the null but forms it by plain residual permutation rather than sign flipping.
Beating the baselines means two things at once. On synthetic benchmarks, spurious admissions must be at or below DP-Forests and recall of true 2- and 3-way sets must be above GA²M and SIAN. On real data, held-out error must be within noise of the best interpretable baseline or better, using equal or fewer terms.
Why not just…?
Why not just use held-out validation to decide which forest-proposed terms to keep? This is the most serious threat. The top-$k$ ablation implements it exactly. Validation loss rewards any term that lowers error, including spurious terms that soak up misfit from main effects on correlated features. Those are precisely the terms an auditor should not see. If the top-$k$ variant matches our spurious-admission rate on Vo-Pati no-interaction and on correlated-feature variants of Friedman, then the null adds nothing and we will report that.
Why not just extend FAST to triples with a heredity rule? It is cheap and may match us on data where 3-way effects have pairwise shadows. We will run it with heredity as the baseline. The decisive test uses LSS-style synthetic functions with a triple conjunction whose pairwise projections are weak, together with Madelon, where ANOVA-BART reports that additive neural models fail. If FAST-with-heredity recovers those triples, our heredity-free proposer is unnecessary.
Why not just rank interactions with SHAP interaction values from XGBoost? SHAP-style summaries can cancel or mask interactions (Beyond Additive Decompositions), and they carry no calibrated threshold. We will nonetheless run them as a ranker feeding the same null test. If they pass as many true sets with fewer false ones, the forest's contribution is only as a candidate generator, and we would swap it out.
Why not just stay with pairs? SIAN reports that a few 2D terms explain 84% of variance on Bike Sharing, so on many datasets pairs may suffice. That is acceptable: the procedure admits no triples when none pass the null. Its claim on real data rests on calibrated admission rather than a guaranteed accuracy gain.
Proposed Experiments and Decisive Checks
The synthetic suite comprises Vo-Pati with and without interactions, the Friedman function, and LSS Boolean conjunctions of order two and three. Each is rerun with input correlation raised from 0 to 0.9 and with sample sizes from 1,000 to 100,000. We measure the false-admission rate against the nominal $\alpha$, recall of true sets, and Jaccard stability of admitted sets across bootstrap resamples.
The first decisive check is on the correlated purely additive designs. The null-calibrated method must stay near $\alpha$, while raw DWP ranking and the LSSFind threshold admit terms at a rate that grows with correlation. If this fails, the central claim fails.
The second decisive check concerns $m_{try}$. Swept across its range, the admitted sets should stay stable under calibration and should not under the fixed threshold.
The real-data suite is Boston Housing (checking whether the DIS × LSTAT interaction reported by DP-Forests is admitted), California Housing, Bike Sharing, socmob with and without a log link, MIMIC-II and MIMIC-III for AUC, and Higgs for scale. On each we report held-out RMSE or AUC, term count, runtime, and the admitted sets.
On socmob we expect to trail the product model's reported 9.48 on the raw scale. If a log link does not bring us near it, we will report that as a limit of additive interaction models rather than tune around it.
References
- Neonatal apnea and hypopnea prediction in infants with Robin sequence with neural additive models for time series
- Interaction Detection with Bayesian Decision Tree Ensembles
- Provable Boolean interaction recovery from tree ensemble obtained via random forests
- Accurate intelligible models with pairwise interactions
- Sparse Interaction Additive Networks via Feature Interaction Detection and Sparse Selection
- Nonparametric sparsity and regularization
- Bayesian Additive Regression Trees for functional ANOVA model
- NODE-GAM: Neural Generalized Additive Model for Interpretable Deep Learning
- Beyond Additive Decompositions: Interpretability Through Separability
- Neural Additive Experts: Context-Gated Experts for Controllable Model Additivity