acceptodds
Under review as a conference paper at ICLR 2027

SPECTRA: Sparse Prototype-Encoded Classification Trees with Routing Attribution

Abstract

Differentiable decision trees aim to combine the transparent structure of trees with the optimization flexibility of neural models. However, the discrete paths presented as explanations do not always reflect the soft routing computation that actually produces a prediction. We introduce SPECTRA, a sparse multiway classification tree that learns prototype-based branches, structural gates, and leaf predictors through differentiable optimization. Its central mechanism is routing attribution, which expresses each branch decision as a combination of feature-level evidence and an explicit structural-gate contribution. Based on the model’s native routing flow, SPECTRA constructs an executable dominant path that identifies the most influential features, quantifies their support for competing branches, and connects local decisions to the final prediction without fitting a post-hoc surrogate. This formulation makes the learned tree structure, routing process, and explanation share the same computational semantics. SPECTRA therefore provides a unified framework for predictive modeling and auditable decision analysis on tabular data, enabling its behavior to be examined in terms of predictive quality, explanation fidelity, structural sparsity, stability, and computational efficiency.

Primary area:
interpretability and explainable AI
Keywords:
Differentiable Decision Trees, Explainable Artificial Intelligence, Auditable Routing, Explanation Fidelity, Tabular Learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_gbebOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors target a critical flaw in the interpretability of soft decision trees. Because soft trees route inputs through all branches at the same time and use structural gates to prune redundant paths, extracting a single hard decision path for human inspection is often unfaithful to the true model behavior. Specifically, selecting a path based on raw routing scores or greedy top-down steps can pick a branch that was actually zeroed out by a structural gate, or miss the globally optimal path because downstream softmax normalizers vary across subtrees. Furthermore, standard soft trees mix input feature effects with learned gate weights and cannot tell when a single path is enough to represent the full weighted mixture of leaves

To solve these problems, the authors propose SPECTRA (soft decision tree) where the explanation comes directly from the exact same math that calculates the prediction. At each branch, the tree learns a simple min–max range for every feature, called a prototype, which gives a clear score for how much each feature helps the routing decision. By taking the logarithm of the branch probabilities and gates, multiplying weights down a path turns into simple addition, allowing a bottom-up search algorithm to always find the exact path with the largest real weight. This same math also splits the difference between the best path and the second-best path into three exact parts: input feature scores, structural gate weights, and node normalization. In addition, the model checks how much of the total weight is held by the top path and uses a validation test to decide when showing one path is safe and when it must show multiple paths instead.

Across 69 tabular datasets, SPECTRA also achieves higher accuracy than standard single decision trees like CART, especially when trees are kept very small with four leaves or fewer so that humans can easily read them.

Critical strengths

Authors target an important issue of soft tree interpretability and achieve comparable (or better) performance in 1 tree setting

Critical weaknesses
  • In standard univariate trees a root-to-leaf path of 2 checks at most 2 featutes, while in soft trees, in general without hardening, all D features are used. Furthermore, the authors admit that SPECTRA's prototype margin spreads 90% of its mass across 6.94 features on average (compared to 2.26 for CART). The node boundary is closer to RBF or shallow MLP with misture of experts rather than standard trees (either univariate or oblique). In such a setting, interpretability is much lower.
  • Across the 69 tasks, the "dominant path" carries an average of only 23.9% of the surviving flow and agrees with the soft tree's actual prediction only 71.74% of the time
  • When L=2L=2 and M=18M=18, SPECTRA is not even a multi-level tree! It is literally a flat, 1-layer soft Mixture-of-Experts (or RBF classifier) with 18 prototype experts at the root node
  • Comparison with univariate trees (CART, MurTree) is unfair because SPECTRA has way more parameters; it should be compared with other soft/multivariate trees. When compared as a single tree at 8 and 16 leaves, SPECTRA loses to CART
Questions for authors

Please address weaknesses

Reviewer_xtmyOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The authors propose SPECTRA, a soft tree that pairs prototype matching with scaled softmax routing and multiplicative hard-concrete gates. Because the model's output is a flow-weighted mixture over all leaves, the paper asks when and how a single root-to-leaf path can faithfully explain the full soft prediction.

Using a simple log-space reparameterization of path flow, the authors extract the maximum-flow path via bottom-up DP and decompose pairwise path margins into feature, gate, and normalizer contributions. To bridge the gap between single-path explanations and the full mixture, they evaluate flow concentration, top-k path reconstruction, and a calibrated rule for explanations.

Critical strengths

Clear question. The authors intended to examine the attribution of the Dominant path / soft-prediction reconstruction, and explore a mechanism which they can attribute the soft trees natively.

Easy-to-follow representation the figures make the work easy to follow.

Critical weaknesses

Performance seems weak Not matching trivial baselines like random-forest may render the proposed method not useful in practice.

The main technical contribution is shallow. Proposition 1 follows directly from taking logs of a product. I failed to see the novelty in this work.

Questions for authors

Some clarifications on technical contribution and main novelty would be greatly appreciated.

Reviewer_r8VoOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

Given that differentiable decision trees assign mass to multiple leaves, determining a path to provide a faithful explanation becomes a challenge. This paper proposes a solution by devising a score for each root-to-leaf path such that the highest-scoring path is also the highest-flow path. For an input x an overlap ratio is computed between the n-th node, k-th branch's prototype and the datapoint. This Tn,kT_{n,k} together with a learned scaling factor ana_{n} is used to compute routing probabilities. In addition, each branch is assigned a gate zn,kz_{n,k} that is learned during training and can suppress less useful branches. The overlap score, scaling factor, and gate are combined to form a branch score, and these scores are accumulated from the root to the leaf to produce a path score. The authors show that this path score is exactly equal to the logarithm of the path flow, allowing paths to be ranked according to their contribution to the model's prediction. For a given input, the paths are sorted in descending order of their scores to obtain the top-k dominant explanation paths. The paper further introduces metrics to quantify how much of the model's prediction is captured by the top-k paths, providing a measure of explanation fidelity in soft decision trees.

Critical strengths
  1. This paper provides a solution to a problem that many soft-tree methods struggle with, whether the displayed explanation actually corresponds to the model's computation.
  2. The dominant-path formulation is elegant and mathematically justified. The paper also provides useful diagnostics such as flow concentration and top-k path reconstruction, which help quantify how faithfully a short explanation captures the behavior of the full model.
  3. The prototype routing is more interpretable and provides feature-level evidence that can be traced back to learnt prototypes.
Critical weaknesses
  1. While the prototype component is interpretable, the gates are essentially learnt optimization variables trained using an L0 regularized objective. A gate can suppress or favor a branch, but its value does not have a particularly intuitive semantic interpretation beyond "the optimizer found this branch useful." Consequently, part of the routing decision comes from quantities that are less interpretable than the prototypes themselves.

  2. The prototypes are also learnt and we don't know if they can be assigned values that are not actually feasible. For example, Sex is Male, age is 8 and they are pregnant. Or A male of age 15 has a mortgage of 1 million and an annual income of 40k dollars etc. Can you show the prototypes used across various branches for 1 or 2 datasets? How are categorical values handled?

  3. The paper reports the fidelity of top-k paths to the full model but does not directly evaluate the standalone test accuracy of the extracted rule set consisting only of those paths. This makes it difficult to judge the efficacy of the dominant paths as an interpretable model.

Questions for authors
  1. Have you considered interpretable alternatives to gates? Some examples that I can think of are (1) You prune the branches by thresholding on their routing scores (2) similar to EBM, remove a branch, see the increase in validation loss and quantify the contribution based on this etc.

Comments on the paper

1 comment
Area_Chair_BrM8OpenReview ↗

Discussion?

Dear authors,

if you want to engage in a discussion, the time is now. The forum will close to the authors on November 18, 2026: https://iclr.cc/Conferences/2027/Dates

Without a rebuttal, there the reviewers are unlikely to change their opinion, which would lead to rejection, most likely.

Sincerely Your AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.