- Rating
- 1
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 1
- Soundness
- 1
Read review
Paper summary
The paper introduces a posthoc XAI method to identify global higher-order feature interactions. It uses pairwise Friedman H-statistics to measure interaction strengths and Leiden clustering with a Constant Potts Model (CPM) objective to identify communities of interacting features. These communities are then ranked based on their importance using permutation importance. The authors empirically compare their method with SHAP, pairwise Shapley interactions, and Archipelago.
Critical strengths
the general goal of finding the global higher-order interaction structure is interesting
Critical weaknesses
I do not think the the methodological approach to find global higher-order interactions is sound and the respective empirical validation is also not convincing as elaborated below.
1. Methodological weaknesses: SINC aims to identify higher-order feature interactions based solely on pairwise H-statistics. However, it is unclear why communities based on a measure that quantifies only pure pairwise interactions that does not account for higher order interactions should be able to identify higher-order interactions. For example, contains only pairwise interactions, but SINC may group all three features into one community implying a three-way interaction between them. Conversely, for with independent, zero-mean features, all pairwise partial dependence functions vanish, although a three-way interaction exists. SINC would therefore not be able to identify this interaction from its pairwise graph. Furthermore, the H-statistic interaction values are normalized by the pairwise PDPs which differ for different interaction pairs and thus the interaction values measure the percentage of pairwise interaction compared to the two respective main effects they account for within the pairwise PDP. Thus, the H-Statistic values are not comparable as absolute interaction values which makes it unclear how their values should be compared and combined as edge weights in the CPM objective. The paper does not justify why this clustering procedure based on the H-Statistic values should recover higher-order interactions rather than simply connected groups of pairwise interactions. It is also unclear why Friedman’s H-statistic was chosen over alternative measures such as ALE-based interactions [1], particularly given the extrapolation problems of partial dependence when features are correlated.
2. Experimental weaknesses: The experimental evaluation does not sufficiently address these methodological concerns and raises additional questions, such as why small nonzero SHAP interaction scores are highlighted as a limitation, whereas SINC explicitly thresholds small interaction values. The comparison with SHAP itself is not very meaningful since it is an aggregation over all weighted effects of a feature not decomposing it into interaction effects. Only adding pairwise Shapley interactions and saying that the combination is more complicated then the proposed approach is in my opinion no real baseline. Particularly since higher-order Shapley interaction decompositions already exist [e.g., 2–4]. It is unclear why these methods were not considered. The comparison with Archipelago could also benefit from a clearer explanation of the methodological differences. In addition, the synthetic experiments do not investigate cases in which pairwise interactions are insufficient to identify the true higher-order interaction structure, nor do they consider correlated features. These experiments would be particularly relevant for assessing the fundamental assumptions of the proposed method -- particular based on the pairwise H-Statistic.
[1] Apley & Zhu (2020). Visualizing the Effects of Predictor Variables in Black Box Supervised Learning Models. Journal of the Royal Statistical Society: Series B.
[2] Fumagalli et al. (2023). SHAP-IQ: Unified Approximation of any-order Shapley Interactions. NeurIPS.
[3] Muschalik et al. (2024). shapiq: Shapley Interactions for Machine Learning. NeurIPS.
[4] Bordt & von Luxburg (2023). From Shapley Values to Generalized Additive Models and back. AISTATS.
Questions for authors
see weaknesses
4 comments
Response to Reviewer dJQh (2/2)
"The comparison with SHAP itself is not very meaningful [...] higher-order Shapley interaction decompositions already exist [2–4]. It is unclear why these methods were not considered."
We thank the reviewer for this comment and for the suggestion to include higher-order Shapley interactions as a baseline. We agree that pairwise SHAP interactions alone are not a sufficient baseline. We keep SHAP in the comparison because it is the most widely used explanation method in practice and therefore a natural reference point. Following the reviewer's suggestion, we now also compare SINC with SHAP-IQ, computed with the shapiq library for a global, loss-based game up to order five, and discuss Shapley interaction indices in the related work. We report k-SII, as FSII, STII and Möbius interactions yield the same recovered interaction structure in our experiments.
On the synthetic benchmark, SINC recovers 72.0% of the ground-truth terms on average, compared with 37.7% for SHAP-IQ. SHAP-IQ assigns the effect of each feature set to its pure interaction component, and since the higher-order components of most benchmark terms are small compared with the main effects of their features, it recovers only one higher-order term. On two of the three real-world datasets Califiorna Housing and Bike Sharing, SHAP-IQ identifies the same dominant interactions as SINC, but requires the maximum interaction order to be chosen in advance and scores every feature set up to that order. Since the number of such feature sets grows combinatorially with the number of features and the interaction order, SHAP-IQ quickly becomes computationally prohibitive. On HIGGS, with 28 features, there are more than 120,000 feature sets of order at most five, so that SHAP-IQ could not be computed up to this order, and with a maximum order of three, every feature set contributes less than 2% of the total, so that no grouping of the features emerges.
At the same time, SHAP-IQ has a clear advantage that we now report explicitly. Since it scores each feature set separately, it can distinguish several pairwise interactions from a single higher-order interaction, which SINC cannot (Appendix I). We therefore see the two methods as complementary, with SHAP-IQ decomposing interactions by order and SINC summarizing them as global communities of interacting features.
"the synthetic experiments do not investigate cases in which pairwise interactions are insufficient to identify the true higher-order interaction structure, nor do they consider correlated features"
We fully agree with the comment. For the first point, we have added two functions, and , which produce the same pairwise interaction network but differ in their higher-order structure (Appendix I). SINC returns the community for both, which is correct only for the latter, while SHAP-IQ distinguishes the two. We now state this explicitly as a limitation. Regarding correlated features, we acknowledge that the synthetic experiments use independent inputs. The real-world datasets contain strongly correlated features, for example latitude and longitude in California Housing and the reconstructed masses in HIGGS. We now discuss in Section 5 how SINC's use of partial dependence and permutation can lead to extrapolation for correlated features, and we point to conditional alternatives such as ALE-based interactions.
I thank the authors for their detailed responses to the rebuttal and their work on the revised manuscript.
Since the authors changed the core of the methodology, the paper needs a major revision, which is not possible within this short rebuttal time. Also, I would highly recommend to add a background section that introduces the clique detection approach and explaining the intuition behind its usefulness in their approach in more detail.
I am still not convinced by scaling the H-Statistic value by the pairwise PD and thresholding based on that. I think this is not solved by the clique detection, since you are still considering the pairs that fullfill H_ij > thau. Therefore, you might not consider a feature pairs with a high interaction only because they also have a high main effect, but you consider very small (maybe spurious) interactions if they have no main effects since even very small effects then always have an H-Statistic value of 1. Therefore, accounting for the magnitude afterwards via permutation importance does not work, since you filtered out potentitally importance one already in the step before.
It is also not clear to me, how Shapley-based importance on the communities helps to mitigate the problem of features appearing in multiple communities (in non-disjoint cases) not be counted multiple times. If communities are treated as players, then the interactions between communities are fairly distributed among those communities, but this does not generally add up on a feature level or account for the overlap. This should be clarified by the authors and stated more formally.
I appreciate the additional comparisons with SHAPIQ, however, the repository was not updated, which makes it hard to verify those results. If calculated exactly, SHAPIQ provides the decomposition into pure interaction effects. The experiments for higher-order interactions are all done based on an MLP for which the authors do not know the learned interaction effects, but making claims about the recovery of interactions of the true-data generating function. To actually understand interaction recovery, the experiments should additionally be done based on the oracle itself.
I thank the authors again for their efforts, but I do not think that the paper is ready for acceptance, therefore I remain my score.
We thank the reviewer for the comment and the continued engagement with our work. We have updated the repository to include the current methodology and the comparison with SHAP-IQ. We would also like to clarify that for the synthetic benchmark functions F1–F10, Table 5, includes the results of SINC and the baseline methods on the oracle function, shown in parentheses, as requested by the reviewer. On the oracle, SINC recovers 83.3% of the ground-truth terms, compared with 46.1% for Archipelago and 37.7% for SHAP-IQ.
Response to Reviewer dJQh (1/2)
We thank the reviewer for the detailed review and the helpful references. Due to the character limit, our response is split into two parts.
We agree with the examples provided. In the revised manuscript, SINC is now presented as a method for identifying global communities of mutually interacting features rather than higher-order interactions. To further strengthen the method, we have also replaced the Leiden algorithm with maximal clique detection, so that every pair within a community must interact. While this does not establish that a community corresponds to a higher-order interaction, it yields communities that are more closely aligned with the observed interaction structure. In particular, it addresses the reviewer's question of why features should be grouped rather than "simply connected groups of pairwise interactions", since features are no longer grouped through indirect chains of interactions but instead must be directly connected to every other feature in the community.
We partly agree. The H-statistic measures the fraction of a feature pair's joint effect that is due to their interaction, so its values are comparable across pairs as proportions, but not as absolute interaction strengths. Consequently, a pair with a relatively weak joint effect can still obtain a high H-statistic value. In the revised manuscript, the concern is mitigated by replacing the CPM-based community detection with maximal clique detection. As a result, edge weights no longer contribute to an objective function and are only used to decide whether two features interact. The absolute relevance of a detected community is assessed separately by the impact weighting. Communities composed of features with limited predictive relevance therefore receive low importance scores, even if their pairwise H-statistics are high. We now discuss this distinction explicitly in the manuscript.
We chose the H-statistic because it provides a single, global interaction strength for each feature pair with an established and efficient approximation, which is what the construction of the interaction network requires. We agree, however, that partial dependence varies features independently and can therefore evaluate the model on unrealistic feature combinations when features are correlated, whereas ALE-based interactions avoid this extrapolation by only considering local changes within the data distribution. The two approaches thus describe different aspects of the model, with partial dependence capturing how the model itself responds to its features and ALE capturing its behavior on realistic inputs. We now discuss this trade-off in the limitations. Since the interaction measure is a modular component of SINC, an ALE-based interaction network could replace the H-statistic without changing the remaining stages, which we consider a promising direction for future work.
We agree that this was inconsistent and thank the reviewer for pointing it out. We have removed the remark.