Organizations: Department of Physics, University of California, Santa Barbara · Department of Electrical and Computer Engineering, University of California, Santa Barbara
Predicting outcomes from spatiotemporal data, such as assessing Parkinson's severity from human motion, often requires explainability in addition to accuracy. Practitioners need to know what drove the prediction, whether it be spatial (a body joint), temporal (a specific window), or both. Shapley values are attributions computed by treating input features as players in a cooperative game and measuring how the model output changes as coalitions of players are removed. Computing them involves choosing what counts as players, how absent players are imputed, and how the game is evaluated. How these choices affect explanation quality is poorly understood for spatiotemporal data, and real data carries no ground-truth attributions to check against. We present MotionBench-XAI, a benchmark built on i) five synthetic spatiotemporal datasets with exact ground-truth Shapley values, ii) three real datasets across clinical gait, electrocardiogram, and audio, iii) ten attribution methods across three player sets, and iv) three trained architectures. We show that grouping features into players, across time, space, or both, matters more than the choice of imputer. Generative imputers improve explanations only when their conditional models are accurate. On real data, the standard proxy metrics select contradictory best methods. We distill these findings into practical insights and provide an open-source pipeline.
Figures & tables
Figure 1: MotionBench-XAI evaluates spatiotemporal Shapley value attributions as a feature-based explainability method. Datasets (first column): Five synthetic spatiotemporal datasets paired with ground-truth Shapley values, and three real-world datasets. Players P (second column): Input features partitioned spatially ( Pjoint ), temporally ( Ptemp ), or across a spatiotemporal grid ( Pcell ). Imputation methods (third column): Strategies for replacing absent players via off-manifold baselines, on-manifold generative models, or temporal SHAP variants. Metrics (fourth column): Evaluation via exact attribution error and linear agreement error for synthetic cases, or proxy metrics (faithfulness, PlayerAOPC) for real-world cases where ground truth is absent.
Figure 2: What can a player be? The three player sets, shown as masking on a skeleton sequence: a time window ( Ptemp ), a spatial joint ( Pjoint ), or a joint × window cell ( Pcell ).
Study
Synthetic ground truth
Temporal data
Conditional imputers
Multiple player sets
BAM ( Yang and Kim, 2019 )
✓
–
–
–
CLEVR-XAI ( Arras et al., 2022 )
✓
–
–
–
Olsen et al. (2024)
✓
–
✓
–
XTSC-Bench ( Höllig et al., 2023 )
✓
✓
–
–
xaitimesynth ( Baer, 2026 )
✓
✓
–
–
SHAPformer ( Hertel et al., 2026 )
✓
✓
–
✓
Table 1: Coverage of MotionBench-XAI’s evaluation dimensions. Synthetic ground truth provides known informative features or numerical attribution targets. Conditional imputers fill missing inputs using the observed inputs and are evaluated within explanation methods. Multiple player sets means comparing different groupings of input features. ✓ indicates a reported component and – an unreported one.
Table 2: The five synthetic datasets. All are tensors x∈RJ×3×T . The columns give each dataset’s spatial coupling, temporal correlation, and marginal distribution, with the full construction in Appendix B . K is the number of temporal windows and M=J×K the number of joint × window cells. Each dataset has N=200 evaluation sequences.
Method
gauss
skeleton
burr
gait
skel+gait
Standard imputers
KS-Marginal
0.36 ± 0.04
0.41 ± 0.09
0.28 ± 0.04
0.21 ± 0.08
0.16 ± 0.07
WindowSHAP-S
0.29 ± 0.06
0.50 ± 0.13
0.30 ± 0.01
0.44 ± 0.04
0.54 ± 0.13
WindowSHAP-D
0.59 ± 0.09
0.71 ± 0.19
0.54 ± 0.12
0.86 ± 0.14
0.80 ± 0.18
TimeSHAP
0.35 ± 0.10
0.37 ± 0.08
0.37 ± 0.15
0.43 ± 0.27
0.50 ± 0.16
KS-Empirical
0.51 ± 0.03
0.47 ± 0.09
0.48 ± 0.03
0.31 ± 0.05
0.28 ± 0.09
Generative imputers
KS-VAEAC
0.12 ± 0.03
0.16 ± 0.05
0.10 ± 0.02
0.76 ± 0.08
0.87 ± 0.04
Table 3: Synthetic results, temporal players ( K uniform windows). EC3 = one minus the Pearson correlation between estimated and ground-truth credits (lower is better), mean ± spread across the classifier architectures that perform sufficiently well (Appendix F ). Each method is graded against the exact ground truth of its own imputer family (game-matched). Gray text indicates periodic datasets (Result 1); they are excluded from the per-column best. EC1 is reported in Appendix C .
Symbol
Meaning
Data
x∈RJ×F×T
instance: J spatial units, F coords, T frames
f
trained classifier being explained
Players
P , M
player set (a partition of x ), number of players
S⊆P
coalition of visible players
Table 6: Notation. The four benchmark axes and their symbols.
Appendix figures & tables12 assets
Supplementary material from the paper’s appendix.
Appendix
Method
gauss
skeleton
burr
gait
skel+gait
Standard imputers, graded against the marginal ground truth ϕ⋆−mar
KS-Zero
0.000 ± 0.000
0.000 ± 0.000
0.000 ± 0.000
0.000 ± 0.000
0.000 ± 0.000
KS-Mean
0.016 ± 0.008
0.017 ± 0.011
0.009 ± 0.004
0.026 ± 0.025
0.019 ± 0.001
KS-Marginal
0.115 ± 0.052
0.105 ± 0.064
0.096 ± 0.044
0.085 ± 0.081
0.102 ± 0.049
WindowSHAP-S
0.074 ± 0.023
0.073 ± 0.037
0.073 ± 0.025
0.085 ± 0.079
0.161 ± 0.010
WindowSHAP-D
0.099 ± 0.034
0.082 ± 0.045
0.091 ± 0.031
0.111 ± 0.104
0.190 ± 0.028
Appendix
Table 7: EC1, temporal players (companion to Table 3 ): mean absolute attribution error against the exact game-matched ground truth, mean ± spread across the classifier architectures that pass the accuracy criterion (Appendix F ). EC1 carries the attribution scale of each dataset, so values compare within a column, not across columns. Gray as in the main table.
Method
gauss
skeleton
burr
gait
skel+gait
Standard imputers, graded against the marginal ground truth ϕ⋆−mar
KS-Zero
0.000 ± 0.000
0.001 ± 0.001
0.000 ± 0.000
0.002 ± 0.001
0.002 ± 0.001
KS-Mean
0.003 ± 0.002
0.002 ± 0.001
0.002 ± 0.001
0.006 ± 0.006
0.004 ± 0.001
KS-Marginal
0.093 ± 0.042
0.030 ± 0.018
0.099 ± 0.045
0.024 ± 0.023
0.027 ± 0.017
KS-Empirical
0.127 ± 0.053
0.042 ± 0.023
0.132 ± 0.058
0.034 ± 0.032
0.042 ± 0.024
Generative imputers, graded against the conditional ground truth ϕ⋆−cond
Appendix
Table 8: EC1, spatial players (companion to Table 4 ), mean ± spread across the classifier architectures that pass the accuracy criterion (Appendix F ), game-matched grading.
Method
gauss
skeleton
burr
gait
skel+gait
Standard imputers, graded against the marginal ground truth ϕ⋆−mar
KS-Zero
0.008 ± 0.005
0.005 ± 0.004
0.012 ± 0.007
0.006 ± 0.006
0.007 ± 0.006
KS-Mean
0.010 ± 0.006
0.006 ± 0.004
0.013 ± 0.008
0.008 ± 0.008
0.008 ± 0.005
KS-Marginal
0.028 ± 0.014
0.016 ± 0.010
0.033 ± 0.016
0.013 ± 0.012
0.015 ± 0.009
KS-Empirical
0.045 ± 0.017
0.021 ± 0.012
0.046 ± 0.019
0.017 ± 0.016
0.023 ± 0.010
Generative imputers, graded against the conditional ground truth ϕ⋆−cond
Appendix
Table 9: EC1, spatiotemporal players (companion to Table 4 ), mean ± spread across the classifier architectures that pass the accuracy criterion (Appendix F ), game-matched grading. Gray as in the main tables.
Method
gauss
skeleton
burr
gait
skel+gait
Standard imputers, graded against the conditional ground truth
KS-Zero
0.25 ± 0.03
0.28 ± 0.09
0.23 ± 0.04
0.87 ± 0.11
0.87 ± 0.13
KS-Mean
0.26 ± 0.02
0.29 ± 0.09
0.23 ± 0.03
0.86 ± 0.10
0.86 ± 0.12
KS-Marginal
0.45 ± 0.06
0.52 ± 0.09
0.38 ± 0.03
0.86 ± 0.08
0.84 ± 0.10
WindowSHAP-S
0.46 ± 0.04
0.65 ± 0.11
0.43 ± 0.04
0.90 ± 0.03
0.91 ± 0.03
WindowSHAP-D
0.68 ± 0.06
0.77 ± 0.15
0.61 ± 0.09
0.94 ± 0.06
0.98 ± 0.02
Appendix
Table 10: Cross-graded EC3, temporal players : every method graded against the on-manifold (conditional) ground truth, the convention of prior oracle benchmarks. Comparing the standard-imputer block with its game-matched values in Table 3 isolates the ground-truth divergence Δ(ϕ⋆−mar,ϕ⋆−cond) from estimator error. Gray columns: on periodic data the conditional ground truth is nearly uniform across windows, so no per-column best is marked.
Method
gauss
skeleton
burr
gait
skel+gait
Standard imputers, graded against the conditional ground truth
KS-Zero
0.35 ± 0.19
0.31 ± 0.16
0.35 ± 0.19
0.00 ± 0.00
0.24 ± 0.19
KS-Mean
0.35 ± 0.19
0.31 ± 0.16
0.36 ± 0.19
0.05 ± 0.05
0.24 ± 0.18
KS-Marginal
0.37 ± 0.05
0.27 ± 0.09
0.39 ± 0.06
0.17 ± 0.15
0.22 ± 0.13
KS-Empirical
0.41 ± 0.02
0.29 ± 0.04
0.43 ± 0.04
0.26 ± 0.13
0.28 ± 0.08
Generative imputers, graded against the conditional ground truth (as in the main tables)
Appendix
Table 11: Cross-graded EC3, spatial players : all methods against the on-manifold ground truth (exact conditional target for M≤12 , hi-budget sampled target above).
Method
gauss
skeleton
burr
gait
skel+gait
Standard imputers, graded against the conditional ground truth
KS-Zero
0.31 ± 0.10
0.31 ± 0.10
0.33 ± 0.08
0.55 ± 0.05
0.72 ± 0.03
KS-Mean
0.31 ± 0.10
0.31 ± 0.10
0.33 ± 0.07
0.55 ± 0.04
0.72 ± 0.03
KS-Marginal
0.34 ± 0.06
0.35 ± 0.05
0.39 ± 0.04
0.59 ± 0.06
0.71 ± 0.04
KS-Empirical
0.39 ± 0.01
0.38 ± 0.02
0.45 ± 0.02
0.58 ± 0.07
0.66 ± 0.02
Generative imputers, graded against the conditional ground truth (as in the main tables)
Appendix
Table 12: Cross-graded EC3, spatiotemporal players : all methods against the on-manifold ground truth. At M=68 ( skel+gait cells) the estimating standard imputers’ cross-graded error is three to five times their game-matched error (Table 4 ): at fine granularity the two ground truths assign very different credit across cells.
Dataset
Classifier
KS-Zero
KS-Mean
KS-Marginal
WindowSHAP-S
WindowSHAP-D
TimeSHAP
KS-Empirical
KS-VAEAC
KS-Flow
KS-Gauss
KS-Oracle
gauss
mlp
0.00
0.02
0.30
0.37
0.72
0.23
0.47
0.07
0.31
0.20
0.15
cnn
0.00
0.01
0.36
0.23
0.56
0.36
0.51
0.14
0.45
0.23
0.21
transformer
0.00
0.02
0.41
0.27
0.50
0.47
0.55
0.14
0.47
0.29
0.27
skeleton
mlp
0.00
0.02
0.29
0.68
0.98
0.26
0.35
0.10
0.26
0.19
0.13
cnn
0.00
0.05
0.44
0.39
0.55
0.37
0.51
0.15
0.38
0.28
0.20
transformer
0.00
0.03
0.50
0.43
0.60
0.46
0.56
0.23
0.49
0.38
0.28
Appendix
Table 13: Per-classifier EC3, temporal players (game-matched). The two † rows fail the accuracy criterion (Appendix F ) and are excluded from every aggregated table; they are shown here for completeness. Per-classifier values for the spatial and cell player sets ship with the canonical result files in the code release.
Player set
Dataset
M
Budget
Floor EC3 (cond / marg)
Floor EC1 (cond / marg)
cells
gauss
20
8,192
0.016 / 0.004
0.0068 / 0.0030
cells
skeleton
68
8,192
0.118 / 0.020
0.0128 / 0.0052
cells
skeleton
68
32,768
0.026 / 0.003
0.0053 / 0.0019
cells
burr
25
8,192
0.020 / 0.013
0.0082 / 0.0048
cells
gait
68
8,192
0.075 / 0.053
0.0065 / 0.0060
cells
gait
68
32,768
0.014 / 0.022
0.0028 / 0.0023
Appendix
Table 14: Grading error floors for the sampled high-budget targets: EC between the stored grading target and an independent replicate (same budget, disjoint coalition seed), averaged over the gate-passing classifier architectures. Method-vs-target differences below these levels are not interpretable. Exact-enumeration cells ( M≤12 ) have zero grading error by construction.
Dataset
Classifier
Train
Val.
Test
Gap
Admitted
Synthetic (chance =1/3 ; accuracy bar 0.383 )
gauss
mlp
0.644
0.662
0.590
+0.054
✓
gauss
cnn
0.951
0.892
0.870
+0.081
✓
gauss
transformer
0.990
0.950
0.930
+0.059
✓
skeleton
mlp
0.533
0.542
0.430
+0.103
✓
skeleton
cnn
0.961
0.878
0.865
+0.096
✓
Appendix
Table 15: The classifier accuracy criterion. A classifier’s explanations enter the aggregated tables only if (G1) held-out accuracy ≥ chance +0.05 and (G2) train–test gap ≤0.30 . Synthetic: accuracies on the training / validation / evaluation splits. Real: held-out accuracy pooled over the three data folds (per-fold values in parentheses); the training column reports accuracy on the pooled training material of the cross-validation folds. Take-away: one rule yields every exclusion. POTR fails G1, and the gait and skel+gait CNNs fail G2; every other classifier is admitted.
Dataset
Task
Classifiers (one per fold)
Track
M
Coalitions
CARE-PD
UPDRS-gait severity (3-class)
MotionBERT (A), MotionAGFormer (B);
windows ( K=4 )
4
exact ( 24 )
17 H36M joints, T=80
POTR excluded (Appendix F )
joints
17
sampled, B=2048
joint × window cells
68
sampled, B=2048
PTB-XL
Normal vs. MI (binary)
ResNet1d
windows ( K=4 )
4
exact ( 24 )
12 leads, T=1000 ( 100 Hz)
leads
12
exact ( 212 )
lead × window cells
48
sampled, B=2048
Appendix
Table 16: Real-data protocol. Tasks, classifiers, and the three player-set tracks per dataset. All tracks: 3 folds, 200 sequences per fold ( 200/180/200 on CARE-PD), folds pooled. Sampled designs share the coalition protocol of Appendix A (seed 7919 ). CARE-PD folds are leave-one-subject-out. ESC-50 uses folds 1 – 3 of its 5 stratified folds.
Faithfulness
PlayerAOPC
Method
CARE-PD A M=68
CARE-PD B M=68
PTB-XL M=48
ESC-50 M=16
CARE-PD A M=68
CARE-PD B M=68
PTB-XL M=48
ESC-50 M=16
KS-Zero
0.89 ± 0.03
0.66 ± 0.23
0.74 ± 0.03
0.77 ± 0.01
-0.11 ± 0.25
0.39 ± 0.13
0.39 ± 0.04
0.72 ± 0.00
KS-Mean
0.82 ± 0.05
0.68 ± 0.15
0.74 ± 0.03
0.78 ± 0.01
0.19 ± 0.04
0.33 ± 0.09
0.39 ± 0.04
0.72 ± 0.00
KS-Marginal
0.79 ± 0.03
0.50 ± 0.03
0.57 ± 0.02
0.87 ± 0.01
0.08 ± 0.08
0.19 ± 0.01
0.32 ± 0.02
0.85 ± 0.01
KS-VAEAC
0.39 ± 0.24
0.54 ± 0.10
0.78 ± 0.03
0.88 ± 0.01
0.02 ± 0.04
0.08 ± 0.01
0.30 ± 0.05
0.80 ± 0.01
KS-Flow
0.42 ± 0.16
0.52 ± 0.07
0.61 ± 0.05
0.86 ± 0.01
0.05 ± 0.03
0.12 ± 0.01
0.26 ± 0.03
0.81 ± 0.01
Appendix
Table 17: Real-world results, spatiotemporal players (joint × window / lead × window / band × window cells). Faithfulness ( ↑ ) and PlayerAOPC ( ↑ ), pooled over the three data folds, ± spread across folds; bold: best per column and metric. Protocol identical to Table 5 (coalition budget 2048, N=200 per fold).
Dataset
Imputer
Spatial
Temporal
Cell
CARE-PD
VAEAC
1.00 ± 0.00
0.96 ± 0.12
1.00 ± 0.00
Flow
1.00 ± 0.00
1.00 ± 0.01
1.00 ± 0.00
Donor control
0.97
0.98
0.98
Unconditional
0.99 (all families)
PTB-XL
VAEAC
0.96 ± 0.09
0.05 ± 0.10
0.98 ± 0.07
Flow
0.95 ± 0.12
0.10 ± 0.15
0.99 ± 0.04
Appendix
Table 18: Imputer capability check : hide-one-recover correlation between the imputed conditional mean and the hidden ground truth, per player-set mask family (mean ± spread over masks; 3 folds, 24 sequences per fold, 8 draws). Donor : copy the hidden entries from the most-correlated other unit of the same sequence – the recovery available from raw redundancy alone. Unconditional : imputer run with everything hidden. A learned imputer conditions on the visible context exactly when its recovery clears both controls.
We address the problem of explainability in machine learning models through feature attribution methods. In particular, we consider a variant of Shapley values known as Asymmetric Shapley Values (ASV), which enables the incorporation of causal knowledge into model-agnostic explanations through the use of a causal graph. We show that in certain contexts in which the computation of SHAP is #P-hard, the exact computation of ASV can be done in polynomial time. To extend this algorithmic result, we introduce a notion of equivalence classes over the topological orderings of the underlying causal graph, which is useful to reduce the time to compute ASV. In particular, we present a polynomial-time algorithm (in the number of equivalence classes) to compute it whenever the causal graph is a rooted directed tree. Finally, we develop an algorithm for approximating ASV in arbitrary causal DAGs which relies on a procedure to sample topological orderings uniformly at random. To implement this sampling mechanism we leverage known algorithms as well as simpler alternatives. Our experimental results demonstrate the practical viability of the proposed approach in realistic causal structures.
Ezequiel Companeetz, Santiago Cifuentes, Sergio Abriola
Departamento de Computación, Facultad de Ciencias Exactas y Naturales, UBA · Instituto de Ciencias de la Computación (ICC), CONICET
Machine learning pipelines commonly flatten relational data into single-table representations, discarding structural constraints. Widely used Shapley value-based feature attributions then rely on feature independence, evaluating the model on combinations that could never arise in the underlying data, producing misleading explanations. We propose RelShap, a framework that incorporates relational constraints and data provenance into Shapley value computation, restricting both background data and coalition evaluation to relationally valid configurations. The framework is estimator-agnostic and composes with Kernel SHAP, Monte Carlo, and Leverage SHAP without altering their sampling or weighting properties. Functional dependencies further induce equivalence classes over feature coalitions, which RelShap exploits to reduce runtime without changing Shapley values; we provide a combinatorial characterization of the expected speedup. Experiments across multiple datasets, models, and estimators show that RelShap produces explanations that are more faithful to the data-generating process, correctly identifying the dominant feature in controlled settings where existing methods, including Conditional SHAP and ManifoldShap, do not. Our code is available at: https://github.com/duneag2/relshap.
Seungeun Lee, Joao Fonseca, Julia Stoyanovich
New York University, New York, USA · INESC-ID, Lisbon, Portugal
We introduce Aumann-SHAP, an interaction-aware framework that decomposes counterfactual transitions by restricting the model to a local hypercube connecting baseline and counterfactual features. Each hypercube is discretized into a grid to construct an induced micro-player cooperative game in which elementary grid-step moves become players. Shapley and LES values on this TU-micro-game yield geometry-aware within-pot attributions that converge to the diagonal Aumann--Shapley / Integrated Gradients limit under grid refinement, and recover equal-split Shapley as the degenerate m=1 special case. An exact grid-state closed form gives polynomial-time computation for fixed interaction order. On a synthetic benchmark with known ground truth, equal-split Shapley carries an irreducible bias while Aumann-SHAP converges to the correct decomposition. On German Credit, interaction geometry changes feature priority rankings in 12.3% of instances. On UCI Heart Disease, equal-split misattributes a cholesterol suppressor as a positive contributor, which is a sign error Aumann-SHAP corrects. On MNIST, game-theoretic attribution reaches target confidence with 3.5× fewer edits than magnitude-based ordering, with micro-game Shapley achieving the best efficiency across all budgets.