Interventional TreeSHAP Explained

• Interventional TreeSHAP is a method that computes exact Shapley attributions for tree models by replacing features outside a coalition with baseline values.
• It exploits the inherent decision tree structure to reduce the exponential subset sum to a polynomial-time computation while distinguishing interventional from path-dependent semantics.
• The technique clarifies model-fidelity by emphasizing marginal feature attributions under interventional assumptions, though it may evaluate on implausible hybrid inputs when features are dependent.

Interventional TreeSHAP is the exact tree-specific computation of Shapley attributions under an interventional coalition semantics for decision trees and tree ensembles. In the proof-oriented formulation of "Understanding Interventional TreeSHAP: How and Why it Works," the game is defined for a model $h$, an explicand $\mathbf{x}$, and a baseline point $\mathbf{z}$ by replacing coordinates outside a coalition $S$ with baseline values and evaluating the resulting hybrid input; later work on partial dependence and marginal feature attribution formulates the same interventional family through expectations of the form $v_S(x_S)=\mathbb{E}_{P_X}[m(x_S,X_{\bar S})]$, which those authors explicitly identify with interventional Shapley values (Laberge et al., 2022, Liu et al., 2024, Filom et al., 2023). The central technical fact is that trees turn the exponential subset sum of the Shapley formula into an exact polynomial-time computation by exploiting root-to-leaf path structure, and the subsequent literature divides sharply between methods that preserve this interventional estimand and methods that accelerate or generalize a different, path-dependent tree-specific semantics (Laberge et al., 2022, Yu et al., 2022, Yang, 2021).

1. Semantic definition and estimands

The supplied literature contains two interventional formulations. The baseline-point formulation, used in the proof paper, defines a replace operator

$$r_i^{\mathbf{z}}(S)= \begin{cases} x_i & \text{if } i\in S,\ z_i & \text{otherwise}, \end{cases}$$

and the coalition game

$\nu_{h,\mathbf{x},\mathbf{z}}(S):=h(\replace{S}).$

The corresponding Shapley value is

$$\phi_i(\nu)=\sum_{S\subseteq [d]\setminus \{i\}} W(|S|,d)\big(\nu(S\cup\{i\})-\nu(S)\big),\qquad W(k,d):=\frac{k!(d-k-1)!}{d!}.$$

By the efficiency axiom,

$$\sum_{i=1}^d \phi_i(h,\mathbf{x},\mathbf{z}) = h(\mathbf{x})-h(\mathbf{z}).$$

This is explicitly an interventional semantics because features are independently “switched on” by replacing baseline values $z_i$ with target values $\mathbf{x}$0, irrespective of feature dependence (Laberge et al., 2022).

The background-distribution formulation, emphasized in the partial-dependence literature, defines

$\mathbf{x}$1

with empirical estimator

$\mathbf{x}$2

That literature states directly that SHAP values built from these partial-dependence functions are what others call interventional Shapley values, and it is equally explicit that this is a marginal/interventional expectation rather than a conditional expectation (Liu et al., 2024). The marginal-game paper uses the notation

$\mathbf{x}$3

for the same object and stresses that these attributions are implementation invariant because they depend only on the input-output function $\mathbf{x}$4 and the feature distribution, not on the internal structure of a trained tree (Filom et al., 2023).

A useful contrast is the tree-path-dependent semantics used in much of the TreeSHAP acceleration literature. There, missing features are handled by following both branches with weights derived from training-time cover statistics or branch proportions. That semantics is not the same as the background-distribution interventional estimand, and several later papers treat the distinction as foundational rather than cosmetic (Yang, 2021, Mitchell et al., 2020, Liu et al., 2024).

2. Path-wise reduction of the Shapley sum

The proof-oriented account of Interventional TreeSHAP treats a binary decision tree as a sum over root-to-leaf path functions. If $\mathbf{x}$5 is a maximal path ending at leaf $\mathbf{x}$6 with leaf value $\mathbf{x}$7, then

$\mathbf{x}$8

This decomposition is the basis for the exact reduction: linearity of Shapley values allows one to compute attributions path by path and then sum them (Laberge et al., 2022).

For a fixed path $\mathbf{x}$9, each edge is classified by whether $\mathbf{z}$0 and $\mathbf{z}$1 traverse it. The four edge types are $\mathbf{z}$2, $\mathbf{z}$3, $\mathbf{z}$4, and $\mathbf{z}$5: only $\mathbf{z}$6 follows the edge, only $\mathbf{z}$7 follows the edge, both follow it, or neither follows it. If the path contains no type-$\mathbf{z}$8 edge, then the interventional path game satisfies

$\mathbf{z}$9

Defining

$S$0

the key flow-blocking lemma states that if $S$1 contains no type-$S$2 edge and $S$3, then

$S$4

Thus, for each relevant feature on the path, the Shapley subset sum collapses to a single nonzero coalition (Laberge et al., 2022).

The resulting closed forms are the core of Interventional TreeSHAP: $S$5

$S$6

$S$7

These formulas show why the tree case is tractable: each path behaves like a sparse Boolean conjunction over the path-relevant features rather than as an arbitrary set function (Laberge et al., 2022).

3. Recursive tree algorithm and ensemble decomposition

The recursive Interventional TreeSHAP algorithm maintains exactly two sets along the current root-to-current-node path: $S$8 and $S$9. Those sets summarize all information needed at a leaf. The traversal distinguishes three cases. If $v_S(x_S)=\mathbb{E}_{P_X}[m(x_S,X_{\bar S})]$0 and $v_S(x_S)=\mathbb{E}_{P_X}[m(x_S,X_{\bar S})]$1 go to the same child, the algorithm follows only the shared child because the other branch would create a type-$v_S(x_S)=\mathbb{E}_{P_X}[m(x_S,X_{\bar S})]$2 edge and therefore zero contribution. If $v_S(x_S)=\mathbb{E}_{P_X}[m(x_S,X_{\bar S})]$3 and $v_S(x_S)=\mathbb{E}_{P_X}[m(x_S,X_{\bar S})]$4 split on a feature already seen earlier on the path, only the branch consistent with the earlier role of that feature is explored, preserving $v_S(x_S)=\mathbb{E}_{P_X}[m(x_S,X_{\bar S})]$5. If they split on a new feature, both branches are explored: the $v_S(x_S)=\mathbb{E}_{P_X}[m(x_S,X_{\bar S})]$6-child adds that feature to $v_S(x_S)=\mathbb{E}_{P_X}[m(x_S,X_{\bar S})]$7, and the $v_S(x_S)=\mathbb{E}_{P_X}[m(x_S,X_{\bar S})]$8-child adds it to $v_S(x_S)=\mathbb{E}_{P_X}[m(x_S,X_{\bar S})]$9 (Laberge et al., 2022).

At a leaf with value $$r_i^{\mathbf{z}}(S)= \begin{cases} x_i & \text{if } i\in S,\ z_i & \text{otherwise}, \end{cases}$$0, the current $$r_i^{\mathbf{z}}(S)= \begin{cases} x_i & \text{if } i\in S,\ z_i & \text{otherwise}, \end{cases}$$1 determine the completed path contribution. For each $$r_i^{\mathbf{z}}(S)= \begin{cases} x_i & \text{if } i\in S,\ z_i & \text{otherwise}, \end{cases}$$2,

$$r_i^{\mathbf{z}}(S)= \begin{cases} x_i & \text{if } i\in S,\ z_i & \text{otherwise}, \end{cases}$$3

$$r_i^{\mathbf{z}}(S)= \begin{cases} x_i & \text{if } i\in S,\ z_i & \text{otherwise}, \end{cases}$$4

The ordinary exponential Shapley sum is never evaluated explicitly. The recursion carries the combinatorial summary $$r_i^{\mathbf{z}}(S)= \begin{cases} x_i & \text{if } i\in S,\ z_i & \text{otherwise}, \end{cases}$$5, and the leaf update already includes the correct Shapley subset weights (Laberge et al., 2022).

For the proof-oriented algorithm, if $$r_i^{\mathbf{z}}(S)= \begin{cases} x_i & \text{if } i\in S,\ z_i & \text{otherwise}, \end{cases}$$6 are represented as binary vectors and cardinalities are tracked, the paper states complexity $$r_i^{\mathbf{z}}(S)= \begin{cases} x_i & \text{if } i\in S,\ z_i & \text{otherwise}, \end{cases}$$7, where $$r_i^{\mathbf{z}}(S)= \begin{cases} x_i & \text{if } i\in S,\ z_i & \text{otherwise}, \end{cases}$$8 is the set of internal nodes and $$r_i^{\mathbf{z}}(S)= \begin{cases} x_i & \text{if } i\in S,\ z_i & \text{otherwise}, \end{cases}$$9 the set of leaves. It then notes that the SHAP implementation uses additional optimizations to propagate leaf contributions upward and reduce the complexity further to $\nu_{h,\mathbf{x},\mathbf{z}}(S):=h(\replace{S}).$0, where $\nu_{h,\mathbf{x},\mathbf{z}}(S):=h(\replace{S}).$1 is the set of nodes. For ensembles, linearity gives

$\nu_{h,\mathbf{x},\mathbf{z}}(S):=h(\replace{S}).$2

so one explains each tree and then sums or averages according to the ensemble output rule (Laberge et al., 2022).

The same proof strategy also extends to Shapley-Taylor indices and grouped features. The paper provides explicit formulas for pairwise Shapley-Taylor terms and a grouped-feature construction for one-hot encoding, where all embedded coordinates corresponding to one original feature are switched on or off together (Laberge et al., 2022).

4. Separation from path-dependent TreeSHAP

A central theme of the later literature is that path-dependent TreeSHAP and interventional TreeSHAP are not merely two implementations of the same value function. Path-dependent TreeSHAP defines coalition values recursively from the internal tree structure, using branch weights based on training-time cover counts or empirical routing proportions. Fast TreeSHAP states that for tree-based models the underlying TreeSHAP derivation uses a conditional expectation view operationalized by recursively following the decision path for $\nu_{h,\mathbf{x},\mathbf{z}}(S):=h(\replace{S}).$3 if the split feature is in $\nu_{h,\mathbf{x},\mathbf{z}}(S):=h(\replace{S}).$4, and taking a weighted average of both branches if the split feature is not in $\nu_{h,\mathbf{x},\mathbf{z}}(S):=h(\replace{S}).$5, with training-data proportions as weights (Yang, 2021). GPUTreeShap describes the same semantics as “cover weighting,” where absent features are resolved by taking both branches with probabilities estimated from weighted training instances (Mitchell et al., 2020).

By contrast, the marginal/interventional literature treats the target as

$\nu_{h,\mathbf{x},\mathbf{z}}(S):=h(\replace{S}).$6

or equivalently as the model partial-dependence function $\nu_{h,\mathbf{x},\mathbf{z}}(S):=h(\replace{S}).$7. Those papers stress two consequences. First, marginal/interventional feature attributions are implementation invariant: if two trees compute the same function, their marginal Shapley values coincide (Filom et al., 2023). Second, path-dependent TreeSHAP need not agree with that estimand. One paper presents two statistically similar decision trees that compute the exact same function but for which path-dependent TreeSHAP yields different feature rankings, whereas the marginal Shapley values coincide (Filom et al., 2023). Another proves an “Inconsistency of TreeSHAP-path” proposition: there exist a distribution $\nu_{h,\mathbf{x},\mathbf{z}}(S):=h(\replace{S}).$8, an evaluation point $\nu_{h,\mathbf{x},\mathbf{z}}(S):=h(\replace{S}).$9, and two distinct but prediction-equivalent trees whose path-dependent explanations differ, and the mismatch with the population SHAP value computed via the model PD function persists even as the background sample size tends to infinity (Liu et al., 2024).

This distinction is especially important under feature dependence. The partial-dependence paper states directly that path-dependent TreeSHAP is inconsistent when features are correlated, while FastPD consistently estimates the desired population quantity (Liu et al., 2024). The marginal-game paper likewise argues that path-dependent TreeSHAP is not interventional Shapley and may violate implementation invariance because it depends on tree topology, split order, and empirical branch proportions at nodes (Filom et al., 2023).

5. Algorithmic developments beyond the original proof

Several later papers extend or accelerate exact tree SHAP computation, but most do not directly target the interventional semantics. "Linear TreeShap" introduces a polynomial-based dynamic programming formulation in which subset sums are encoded in summary polynomials and exact single-tree SHAP computation improves from $$\phi_i(\nu)=\sum_{S\subseteq [d]\setminus \{i\}} W(|S|,d)\big(\nu(S\cup\{i\})-\nu(S)\big),\qquad W(k,d):=\frac{k!(d-k-1)!}{d!}.$$0 to $$\phi_i(\nu)=\sum_{S\subseteq [d]\setminus \{i\}} W(|S|,d)\big(\nu(S\cup\{i\})-\nu(S)\big),\qquad W(k,d):=\frac{k!(d-k-1)!}{d!}.$$1 with the same $$\phi_i(\nu)=\sum_{S\subseteq [d]\setminus \{i\}} W(|S|,d)\big(\nu(S\cup\{i\})-\nu(S)\big),\qquad W(k,d):=\frac{k!(d-k-1)!}{d!}.$$2 memory. However, the paper is explicit that its semantics are induced by branch-weighted missing-feature traversal and not by an explicitly interventional expectation $$\phi_i(\nu)=\sum_{S\subseteq [d]\setminus \{i\}} W(|S|,d)\big(\nu(S\cup\{i\})-\nu(S)\big),\qquad W(k,d):=\frac{k!(d-k-1)!}{d!}.$$3 or $$\phi_i(\nu)=\sum_{S\subseteq [d]\setminus \{i\}} W(|S|,d)\big(\nu(S\cup\{i\})-\nu(S)\big),\qquad W(k,d):=\frac{k!(d-k-1)!}{d!}.$$4 (Yu et al., 2022). "Fast TreeSHAP" develops Fast TreeSHAP v1 and v2, reports that v1 is 1.5x faster and v2 is 2.5x faster than TreeSHAP, and shows that v2 can be as high as 3x faster for repeated explanations of newly incoming samples, but its derivation is likewise anchored in tree-path branch-cover semantics rather than a separate interventional analysis (Yang, 2021). "GPUTreeShap" reformulates exact TreeShap for SIMD/SIMT execution and reports speedups of up to 19x for SHAP values and up to 340x for SHAP interaction values, again for the cover-weighted path-dependent formulation rather than the background-dataset interventional one (Mitchell et al., 2020).

The interaction literature follows the same pattern. "Beyond TreeSHAP: Efficient Computation of Any-Order Shapley Interactions for Tree Ensembles" presents TreeSHAP-IQ, a polynomial-arithmetic method for exact any-order Shapley interactions in a single recursive traversal, but states explicitly that its formal development uses path dependent feature perturbation and only notes that an interventional version can be computed with the usual added background-sampling cost (Muschalik et al., 2024). By contrast, "Fast Estimation of Partial Dependence Functions using Trees" directly targets the partial-dependence value function

$$\phi_i(\nu)=\sum_{S\subseteq [d]\setminus \{i\}} W(|S|,d)\big(\nu(S\cup\{i\})-\nu(S)\big),\qquad W(k,d):=\frac{k!(d-k-1)!}{d!}.$$5

and says that this is exactly the interventional/marginal SHAP estimand. Its complexity table lists TreeSHAP-int as $$\phi_i(\nu)=\sum_{S\subseteq [d]\setminus \{i\}} W(|S|,d)\big(\nu(S\cup\{i\})-\nu(S)\big),\qquad W(k,d):=\frac{k!(d-k-1)!}{d!}.$$6 and FastPD as $$\phi_i(\nu)=\sum_{S\subseteq [d]\setminus \{i\}} W(|S|,d)\big(\nu(S\cup\{i\})-\nu(S)\big),\qquad W(k,d):=\frac{k!(d-k-1)!}{d!}.$$7, and the paper proves consistency of FastPD for the empirical and population PD targets (Liu et al., 2024).

A separate line of work studies exact marginal feature attributions for tree models without relying on path-dependent recursion. "On marginal feature attributions of tree-based models" argues that marginal/interventional Shapley, Owen, and Banzhaf values are the correct model-faithful objects, identifies a piecewise-constant grid structure for these quantities, and exploits the fact that only a portion of all features appears in a tree from the ensemble. For CatBoost models with oblivious trees, it derives explicit formulas for marginal Shapley values from internal model parameters, with online explanation complexity $$\phi_i(\nu)=\sum_{S\subseteq [d]\setminus \{i\}} W(|S|,d)\big(\nu(S\cup\{i\})-\nu(S)\big),\qquad W(k,d):=\frac{k!(d-k-1)!}{d!}.$$8 after precomputation (Filom et al., 2023).

6. Interpretation, reporting practice, and limitations

Interventional TreeSHAP is most naturally read as a contrastive local explanation. In the baseline-point formulation, $$\phi_i(\nu)=\sum_{S\subseteq [d]\setminus \{i\}} W(|S|,d)\big(\nu(S\cup\{i\})-\nu(S)\big),\qquad W(k,d):=\frac{k!(d-k-1)!}{d!}.$$9 measures the average marginal effect of replacing feature $$\sum_{i=1}^d \phi_i(h,\mathbf{x},\mathbf{z}) = h(\mathbf{x})-h(\mathbf{z}).$$0’s baseline value $$\sum_{i=1}^d \phi_i(h,\mathbf{x},\mathbf{z}) = h(\mathbf{x})-h(\mathbf{z}).$$1 by its value $$\sum_{i=1}^d \phi_i(h,\mathbf{x},\mathbf{z}) = h(\mathbf{x})-h(\mathbf{z}).$$2, averaged over all activation orders, and the attributions sum to $$\sum_{i=1}^d \phi_i(h,\mathbf{x},\mathbf{z}) = h(\mathbf{x})-h(\mathbf{z}).$$3 (Laberge et al., 2022). In the background-distribution formulation, the method explains the trained model by integrating over a background distribution or empirical background dataset, and this remains well-defined even when features are dependent; however, that literature is explicit that such explanations are model-based and not automatically causal or support-respecting for the underlying data-generating process (Liu et al., 2024).

The principal caveat is the interventional assumption itself. Because the semantics break dependence, the method may evaluate the model on unrealistic hybrid inputs. The proof paper highlights this directly: if features are strongly dependent, the interventions may not correspond to plausible data points, and conditional semantics may be preferred when one wants explanations tied to the data distribution rather than to hypothetical feature replacements (Laberge et al., 2022). The partial-dependence paper makes a parallel distinction: interventional SHAP remains the principled PD-based estimand for model explanation under dependence, but not a causal claim about what would happen in the data-generating process without stronger assumptions (Liu et al., 2024).

A second limitation is reporting ambiguity in applied work. An empirical geoscience study on landslide susceptibility uses TreeSHAP as a post hoc explanation and feature-selection tool for an optimized XGBoost classifier, ranks features by mean absolute SHAP value, selects the 9 most significant factors out of 15, and reports an improvement in cross-validation weighted F1 from 94.62% to 95.01% with a final ROC-AUC of 0.97 (Inan et al., 2022). Yet that paper never uses the term “interventional TreeSHAP,” does not specify a background or reference dataset, does not state whether computation was interventional or tree-path-dependent, and does not identify software settings that would distinguish those modes. This is not unusual. A plausible implication is that many applied uses of “TreeSHAP” are semantically underspecified unless the value function, background data, and perturbation mode are reported explicitly.

Taken together, the literature situates Interventional TreeSHAP as a family of exact Shapley algorithms for tree models whose defining property is the interventional coalition semantics rather than tree-path cover weighting. Its mathematical core is the collapse of the path-wise Shapley sum to closed forms indexed by $$\sum_{i=1}^d \phi_i(h,\mathbf{x},\mathbf{z}) = h(\mathbf{x})-h(\mathbf{z}).$$4 and $$\sum_{i=1}^d \phi_i(h,\mathbf{x},\mathbf{z}) = h(\mathbf{x})-h(\mathbf{z}).$$5; its conceptual significance lies in separating model-faithful marginal or baseline-replacement attributions from implementation-dependent path semantics; and its ongoing research frontier concerns efficient computation of the interventional estimand itself, especially under large background sets, higher-order interactions, and alternative tree architectures (Laberge et al., 2022, Liu et al., 2024, Filom et al., 2023).