Partition-Guided Distance Saliency: Bridging Decision and Objective Spaces in Many-Objective Optimization

Abstract: Explainability in Many-Objective Optimization (MaO) is currently hindered by the escalating complexity of the Pareto front, which renders the relationship between high-dimensional decision variables and objective outcomes increasingly opaque. As the number of objectives exceeds the limits of traditional visualization, decision-makers encounter a cognitive drought'' in identifying relevant trade-offs or specifying target regions without a priori knowledge. To bridge this interpretability gap, we introduce the {Partition-Guided Distance Saliency (PGDS)} framework, a novel XAI approach designed for continuous optimization landscapes. Our framework automates the explanation process through a three-stage pipeline that prioritizes geometric intuition over abstract rules. First, we employ a surrogate model that learns how geometric distances in the decision space map to proximity in the objective space. Second, to address the difficulty of manual target selection in high dimensions, the framework automatically partitions the objective landscape into distinct regions and identifies localDominating Points'' to serve as automated targets for improvement. Third, we quantify how sensitive a solution's position is to each decision variable by measuring the distance shifts induced by perturbations to each variable. This allows PGDS to categorize features as either Drivers'' which facilitate convergence toward preferred regions, orBlockers'' which represent geometric constraints hindering further progress. Validation on 10-objective benchmarks and a physics-informed engineering problem (Welded Beam) demonstrates that PGDS provides differentiated, actionable insights that traditional visualization and rule-based XAI methods fail to provide.

• The paper introduces PGDS, a geometry-preserving framework that combines Minimal Learning Machine distance surrogates, KD-Tree target discovery, and RISE-style saliency to explain many-objective optimization solutions.
• The paper demonstrates high surrogate fidelity with R² > 0.99 on DTLZ7 and verifies actionable explanations through a 9.4% improvement and 2.7% degradation in target proximity after changing identified blocker and driver variables.
• The paper shows that variable importance depends on the objective-space region, identifying different drivers and blockers in Welded Beam and 10-objective WFG3 problems while noting open challenges in hyperparameter tuning, baseline comparison, and prescriptive optimization.

Overview

The paper introduces Partition-Guided Distance Saliency (PGDS), a post-hoc explainability framework for many-objective optimization (MaO) that operates directly on continuous decision and objective spaces rather than on discretized or rule-based representations. The work is motivated by a persistent gap in evolutionary multi-objective optimization: while algorithms such as NSGA-III (2606.30836) and MOEA/D produce well-converged Pareto front approximations, decision makers (DMs) lack tools to understand why a solution sits where it does in the objective space, or which decision variables control its position. The authors frame this as a "cognitive drought" in high-dimensional spaces, where visualization techniques such as PCA, t-SNE, and SOM-based projections distort global geometry (2606.30836), and existing XAI-for-optimization methods such as R-XIMO and XLEMOO either require discretization of continuous landscapes or impose a target-selection burden on the DM.

PGDS addresses three research questions: (RQ1) bridging the interpretability gap without losing geometric fidelity; (RQ2) automating discovery of regions of interest; and (RQ3) quantifying sensitivity of objective-space positioning to decision-variable perturbations via a distance-based surrogate. The framework is validated on DTLZ7, the constrained Welded Beam design problem, and a 10-objective WFG3 instance with 20 decision variables.

Methodology

The pipeline has three stages. First, a Minimal Learning Machine (MLM) surrogate is trained on the optimization archive. Rather than regressing raw coordinates, MLM learns a linear map $\Delta_y = D_x B$ between input distance matrices (distances from each solution to $K$ randomly sampled reference points in decision space) and output distance matrices (distances to corresponding reference points in objective space). The regression matrix is solved by ordinary least squares at complexity $\Theta(K^2N)$, and query coordinates are reconstructed via multilateration — analogous to GPS trilateration — using Levenberg–Marquardt. A key property emphasized by the authors is that linearity between distance matrices does not imply linearity between variables and objectives; the formulation sidesteps ill-conditioned coordinate transformations while preserving the manifold structure of the front. The authors claim saliency rankings are stable across random reference-point selections provided $K$ is sufficiently large, citing robustness results for MLM itself rather than presenting their own variance analysis.

Second, automated target discovery uses KD-Tree partitioning of the objective archive into hyper-rectangular leaf blocks, splitting along the highest-variance dimension until a maximum depth or minimum point count is reached. Each block receives a "Dominating Point" defined as the component-wise minimum of its members — a local utopia that serves as an automated explanation target, presented to the DM together with the explicit region constraints (e.g., $f_1 < 0.5 \land f_2 > 0.8$). This directly addresses RQ2 and relieves the DM of manual preference articulation, though the partition depth remains a user-specified parameter.

Third, saliency generation adapts the Distance Explainer methodology (2606.30836) and the RISE masking scheme (2606.30836) to tabular data. Binary masks replace masked variables with the population mean, the MLM predicts the perturbed solution's objective-space coordinates, and the shift in Euclidean distance to the target is recorded over $L$ mask samples. Saliency scores are computed as correlations between variable presence and distance shifts, yielding two categories: Drivers (variables whose preservation keeps the solution near the target) and Blockers (variables whose current values hinder convergence).

Experimental validation

The DTLZ7 walkthrough demonstrates the full user journey on a disconnected, multi-modal front. With $K = N/2$ anchors, the MLM achieves $R^2 > 0.99$ and MSE ≈ 0.0002 across all three objectives, establishing the fidelity prerequisite for reliable explanations. The KD-Tree correctly segments the disconnected front, and the saliency map identifies $x_2$ as the primary driver and $x_1$ as a blocker. These diagnoses are verified causally: increasing $K$0 by 0.18 toward the target configuration improves proximity to the target by 9.4%, while decreasing the driver $K$1 by 0.05 degrades it by 2.7%. This closed-loop verification — applying the suggested delta and measuring actual geometric improvement — is the strongest methodological feature of the evaluation, as it tests actionable content rather than plausibility alone.

The Welded Beam experiment provides physics-informed validation. NSGA-II generates a 200-solution front; a low-cost "trap" solution sits essentially on the buckling constraint boundary ($K$2). Targeting the best-deflection point manually, PGDS assigns beam width $K$3 a normalized blocker saliency of 1.0. Relaxing $K$4 from 0.2363 to 0.2599 moves the critical constraint from −0.0001 to −0.0049, confirming that the framework detected the buckling "wall" induced by $K$5 without access to the governing equations. Notably, the paper concedes that blindly adjusting other variables while holding $K$6 fixed would likely produce infeasible designs — the explanation carries feasibility information that pure objective-space analysis would miss.

The WFG3 experiment scales to 10 objectives and 20 variables, with NSGA-III producing 108 non-dominated solutions after 400 generations. Using only automated KD-Tree targets ("blind targeting"), PGDS distinguishes knee and extreme regions and produces context-dependent saliency: $K$7 is the dominant driver (score −1.00) in the knee region, whereas $K$8 becomes a maximal blocker (+1.00) in the extreme region. The authors interpret these as keystone parameters controlling the degenerate front's curvature, consistent with WFG3's structure. The central claim here is that variable importance is not static but region-dependent — a single global attribution would be misleading, and PGDS's partition-conditioned saliency captures this. The paper also positions traditional parallel coordinate plots as a "failing baseline" since they offer no guidance on where to look or what to adjust.

Limitations and open questions

The paper is explicit about several limitations. No quantitative comparison against rule-based XAI baselines (XLEMOO, R-XIMO) is performed; the authors attribute this to the differing output modalities (logical predicates versus continuous saliency maps) and call for standardized explanation-quality indicators to enable such benchmarking. A systematic sensitivity analysis of the hyperparameters $K$9, $\Theta(K^2N)$0, and $\Theta(K^2N)$1 is deferred, so practitioners currently lack tuning heuristics for varying archive sizes, and the claimed seed-stability of saliency rankings is not empirically demonstrated within this paper. Most substantively, PGDS is diagnostic rather than prescriptive: it identifies influential variables but does not invert the surrogate to compute the decision vector $\Theta(K^2N)$2 reaching a target $\Theta(K^2N)$3, a problem the authors note is ill-posed when $\Theta(K^2N)$4. Validation beyond three benchmark problems, and specifically on industrial problems with heterogeneous constraint structures, remains open.

Conclusion

PGDS contributes a geometry-preserving explainability pipeline for MaO that combines MLM-based distance regression, KD-Tree-driven automated target discovery, and RISE-style tabular saliency. Its empirical support rests on high surrogate fidelity ($\Theta(K^2N)$5), causal verification of driver/blocker hypotheses (9.4% and 2.7% distance effects on DTLZ7; constraint relaxation on Welded Beam), and differentiated region-dependent attributions in a 10-objective setting. The framework's principal open problems are prescriptive inversion of the surrogate, standardized comparison with rule-based XAI, and hyperparameter robustness characterization.