Diffusion Counterfactuals for Manifold-Aware Explanations

• DiffusionCounterfactuals are methods that use diffusion geometry and generative models to produce counterfactual explanations that honor data connectivity.
• They integrate diffusion distances, sparsity constraints, and directional coherence to ensure counterfactuals are both plausible and actionable.
• Empirical evaluations demonstrate improved validity and interpretability across diverse domains such as healthcare, finance, and structured data analysis.

DiffusionCounterfactuals refer to a class of methods that utilize diffusion geometry or diffusion-based generative modeling to generate, evaluate, or explain counterfactuals, particularly in high-dimensional or structured data settings. These methods leverage the transport and connectivity structure induced by diffusion distances or generative diffusion processes to impose manifold-awareness, actionability, and plausibility on counterfactual explanations. The concept has appeared in diverse forms across multiple fields including tabular and structured data, image and graph domains, and personalized counterfactual inference, advancing both interpretability and causal reasoning.

1. Diffusion Distances and Manifold-Aware Counterfactuals

A central insight introduced in "Enhancing Counterfactual Explanation Search with Diffusion Distance and Directional Coherence" is incorporating diffusion distance—a spectral, data-adaptive measure of connectivity—into the optimization objective for counterfactual search (Domnich et al., 2024). The diffusion distance $D_\mathrm{diff}(x, x'; t)$ between points $x$ and $x'$ at diffusion time $t$ is defined via the Zelnik-Manor & Perona self-tuning kernel: $$K_{ij} = \exp\left( -\frac{\|x_i-x_j\|^2}{\sigma_i \sigma_j} \right)$$ where $\sigma_i$ is the local distance to the $k$-th nearest neighbor. After constructing the Markov matrix $P = D^{-1}K$, diffusion coordinates $$\Phi_t(x_i) = (\lambda_1^t\psi_1(i), \ldots, \lambda_L^t\psi_L(i))$$ (where $(\lambda_l, \psi_l)$ are the top $L$ eigenpairs of $P$), yield

$$D_\mathrm{diff}(x_i, x_j; t) = \|\Phi_t(x_i) - \Phi_t(x_j)\|_2$$

By penalizing the diffusion distance between the factual instance and counterfactual proposal, the search is regularized to traverse low-resistance paths on the manifold, avoiding actionless, data-sparse regions.

2. Coherent Directional Counterfactual Explainer (CoDiCE) Methodology

The CoDiCE framework formalizes counterfactual search as minimizing an objective: $$\mathcal{L}(c) = \mathrm{loss}(f(c), y) + \lambda_1 D_\mathrm{diff}(c, x) + \lambda_2 \|c-x\|_0 + \lambda_3 [1 - \mathrm{dcoherence}(c, x)]$$ where the first term enforces prediction-flipping, $D_\mathrm{diff}$ encodes data geometry, sparsity ($\ell_0$) supports actionability, and dcoherence (directional coherence, see below) encourages semantic interpretability.

The search is implemented as a derivative-free genetic algorithm, iteratively evolving offspring in the population based on fitness (negative $\mathcal{L}$), and using precomputed diffusion map distances for efficiency. In mixed-type feature spaces, discrete/categorical attributes are treated using analogous local graph metrics.

3. Directional Coherence: Human-Aligned Edit Semantics

Directional coherence is a regularizer that enforces agreement between the joint direction of feature change in the counterfactual and the model's local, featurewise marginal sensitivity:

• For each feature $i$, set $x_{(i \to c_i)}$ as $x$ with $i$-th entry replaced by $c_i$.
• Compute sign $s_i = \mathrm{sign}(f(x_{(i \to c_i)})_y - f(x)_y)$, representing the model's preference for increasing/decreasing feature $i$ toward $y$.
• The coherence score is

$$\mathrm{dcoherence}(c, x) = \frac{1}{n}\sum_{i=1}^n \mathrm{sgn}[(c_i-x_i)(x'_i-x_i)]$$

penalizing counterfactuals that flip features in directions that decrease the target class probability.

4. Empirical Performance and Ablation Analysis

Experiments confirm that increased diffusion distance regularization (higher $\lambda_1$) yields counterfactuals that stay close to the density manifold, at the expense of sparsity and sometimes coherence. Key findings (Domnich et al., 2024):

• In a diabetes dataset, "diffusion + sparsity" (only $\lambda_1,\lambda_2$ active) produced validity 100%, diffusion ≈1.49, $L_1$ ≈1.01, dcoherence ≈0.56.
• On synthetic "Swiss roll", $L_1$-minimizing (nearest neighbor) counterfactuals cut through low-density; diffusion distance-minimizing CEs respect the manifold.
• On structured and real datasets (Breast Cancer, Adult, COMPAS): CoDiCE achieves best diffusion distance and high validity, with tunable tradeoffs via $\lambda$.

Performance metrics include: validity (% of CEs achieving $f(c)=y$), diffusion distance, $L_1$ (continuous), $L_0$ (categorical), and dcoherence.

5. Computational Complexity and Hyperparameterization

• Kernel construction and spectral decomposition are $O(n^2)$ in data size; truncated spectral methods and approximate nearest neighbor search can mitigate this.
• Key hyperparameters: $k$ (local neighbor count, typically 10–50), $t$ (diffusion time, local/global geometry, typically 1–3), $L$ (number of eigenpairs, spectral gap, 10–50).
• Diffusion distances are precomputed, yielding efficient online queries: only $O$(population size × per-candidate cost) per query.

6. Context, Related Methods, and Theoretical Guarantees

DiffusionCounterfactuals represent a departure from naive $L_p$-based proximity in counterfactual search, grounding plausibility in the global geometry of data space. Unlike approaches relying on binary search or adversarial optimization, the diffusion-based metric incorporates high-order connectivity and promotes traversals through dense, valid data regions, mitigating the risk of generating implausible or unsafe recourses.

From a theoretical perspective, the constructed diffusion kernel is strictly positive semidefinite, and local consistencies for hazard-ratio estimation are provable (in the context of potential outcomes, see (Cloninger, 2016)). Empirically, the diffusion-based approach stabilizes in high dimensions and provides actionable individualized effect estimation in treatment response scenarios.

7. Broader Implications and Extensions

DiffusionCounterfactuals have been extended to handle mixed data types, categorical variables, and generalized to domains beyond tabular, including graphs and structured objects (see (Bechtoldt et al., 20 Nov 2025)), with analogous formulations for structural minimality and class-conditioned generation.

The methodology is directly applicable to high-stakes fields such as healthcare, finance, and recourse generation, where proximity on the data manifold embodies both likelihood and feasibility. Moreover, the separation between metric construction and optimization routine (black-box solver agnosticism) allows further integration with other counterfactual constraints, group fairness regularization, or data augmentation pipelines.

In summary, the DiffusionCounterfactuals paradigm refines the search for feasible, actionable, and interpretable counterfactual explanations through the explicit modeling of data connectivity and local prediction semantics, supporting rigorous deployment and trust in AI-driven decisions (Domnich et al., 2024).