acceptodds
Under review as a conference paper at ICLR 2027

DICTUM: Diffusion Counterfactuals for Tabular Mixed-type Data

Abstract

Counterfactual explanations (CFs) provide actionable recourse by identifying minimal input changes that alter a model’s prediction. Existing methods either optimize a single CF for a fixed criteria or generate multiple CFs from narrow distributions, limiting their ability to adapt to diverse user preferences at inference time. We propose DICTUM, a conditional diffusion framework for tabular data that models a broad distribution over the counterfactual space by training on opposite-class neighbors selected via a Determinantal Point Process (DPP). This diverse training signal, combined with a Transformer-based mixed-type dif- fusion backbone featuring per-column learnable noise schedules, enables DICTUM to generate counterfactual sets with substantially higher distributional coverage than prior generative approaches. At inference time, a branching generation strategy, inspired by SVDD-style guidance, enforces user-specified actionability constraints on the learned distribution, without retraining. Experiments on five tabular benchmarks show that DICTUM achieves competitive or superior performance across validity, proximity, sparsity, plausibility, and diversity metrics. A constraint-strictness sweep and a component ablation further confirm that quality and plausibility are preserved as constraints tighten.

Primary area:
interpretability and explainable AI
Keywords:
counterfactual explanations, tabular data, explainable AI

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_cnxsOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DICTUM, a counterfactual generation method for mixed-type tabular data. It first employs a kk-DPP to select opposite-class targets that balance proximity and diversity, and then trains a conditional diffusion model incorporating branching sampling, constraint projection, and kk-DPP post-processing to generate counterfactual sets.

Critical strengths
  1. The motivation is reasonably sound. The method considers a broader range of alternatives than relying solely on nearest-neighbor training targets.
  2. The results are generally good, with favorable performance in continuous-feature proximity, LOF, and diversity across five datasets.
Critical weaknesses
  1. Lack of novelty. The mixed-type diffusion backbone and per-column noise schedules are largely inherited from TabDiff, the branching guidance draws on SVDD, and DPP is an existing set-selection tool. The contribution appears primarily to be a combination of these existing methods.

  2. DiCE already uses DPP to encourage diversity in counterfactual sets, DiCoFlex already supports diverse counterfactual generation and inference-time constraint adjustment. Therefore, I think, “generating diverse counterfactuals and adapting to user requirements at inference time” does not, on its own, constitute a major innovation. If my understanding is correct, the baselines randomly select valid candidates, whereas DICTUM uses k-DPP post-processing. The same post-processing should be applied to all methods before comparing the generative models.

  3. The coverage and comparison of recent related work are insufficient. For example, both [1] and [2] have studied counterfactual explanations for tabular data, and [2] specifically investigates diffusion-based counterfactual generation for mixed-type tabular data.

  4. Algorithm 1 explicitly filters out candidates whose predictions do not match the target class or that violate constraints before applying k-DPP. Therefore, a final validity close to 1 does not directly demonstrate a high success rate for the original generation process.

[1] Density-Guided Robust Counterfactual Explanations on Tabular Data under Model Multiplicity [2] Tabular Diffusion Counterfactual Explanations

Questions for authors

See Weaknesses.

3 comments

Response to Reviewer cnxs

We thank the reviewer for the feedback.


W1 + W2.1. Novelty

DICTUM is not a backbone swap. Its core idea is to decouple learning a broad counterfactual distribution p(x′|x,y′) from imposing user preferences on it at inference. This needs two things DiCoFlex lacks: training targets covering several recourse strategies, not one neighbourhood (k-DPP vs. kNN), and a sampler with intermediate states, so any black-box constraint can steer generation step by step (the branching noted in S2). A flow maps noise to a sample in one pass: it can only filter finished samples, and its control is limited to constraint types seen in training (no monotonicity). The ablations support combining the two: with our kk-DPP selection given to DiCoFlex (response to 5tWg), DICTUM’s generator alone still yields lower LOF and higher diversity, and step-wise branching meets a feature-change budget that post-hoc filtering cannot enforce.

W2.2. Same post-processing for all methods

The kk-DPP selection is part of DICTUM by design, not an add-on. DICTUM is trained on kk-DPP-curated targets, so it learns a distribution over diverse valid counterfactuals, and the post-hoc selection is the inference counterpart of the same Gower-based kernel. Each baseline is evaluated with the selection of its original paper: DiCE optimises diversity within its own objective.

We applied our kk-DPP post-processing to DiCoFlex and removed it from DICTUM (response to 5tWg). Even without kk-DPP, DICTUM has the best LOF on all five datasets and higher diversity than every baseline except DiCE on Default and GMC, whose proximity is 3–6× worse (Table R1). The gap therefore comes mainly from the generator, not the selection.

Table R1.

MethodProx. ↓Spars. ↓LOF ↓Diversity ↑
Adult
DiCE1.080.140.500.57
CCHVAE0.340.280.400.17
DiCoFlex0.890.290.450.35
DICTUM w/o kk-DPP0.340.290.170.64
DICTUM0.230.270.170.73
Bank
DiCE1.120.090.260.58
CCHVAE0.500.190.130.06
DiCoFlex0.890.130.220.34
DICTUM w/o kk-DPP0.510.190.120.61
DICTUM0.340.170.140.73
Credit Default
DiCE1.180.100.830.80
CCHVAE0.430.380.440.08
DiCoFlex0.600.250.300.34
DICTUM w/o kk-DPP0.350.230.150.63
DICTUM0.260.190.170.75
GMC
DiCE2.330.211.150.83
CCHVAE0.370.320.540.12
DiCoFlex0.360.030.680.52
DICTUM w/o kk-DPP0.370.180.200.77
DICTUM0.250.140.200.91
Lending Club
DiCE2.250.130.840.46
CCHVAE0.570.450.160.07
DiCoFlex0.670.180.100.33
DICTUM w/o kk-DPP0.500.310.070.64
DICTUM0.370.250.070.75

W3. Related work

We thank the reviewer for the references and have added both to §2. [2] (TDCE) guides an unconditional tabular diffusion model with gradients of a differentiable classifier, using a Gumbel-softmax relaxation for categorical features, and produces one counterfactual per query. DICTUM instead learns p(x′∣x,y′)p(\mathbf{x}'\mid\mathbf{x},y') from diverse targets, needs only the classifier's predictions, and returns a diverse set per query under immutable and monotone constraints. [1] (DensityFlow) trains a density-guided Neural ODE so that single counterfactuals stay valid under model multiplicity, which is complementary to our goal.

W4. Validity and filtering

The "Validity" column measures the returned set, not the sampler. It is the fraction of the k=10k=10 returned counterfactuals classified as y′y', computed after the P=100P=100 candidates per query have been filtered. The filter gives DICTUM no advantage and we claim none. A query whose pool contains no valid candidate is not discarded. It receives invalid candidates, and these count as failures. A value of 1.00 therefore means that, up to rounding, each of the Q=2000Q=2000 queries had at least k=10k=10 valid candidates among its 100.

We report pool validity separately: the fraction of the P=100P=100 raw candidates classified as y′y' before any filtering, for every method and dataset.

Table R2. Pool validity before filtering (P=100P=100, Q=2000Q=2000).

MethodAdultBankDefaultGMCLending
DiCE0.900.910.900.930.77
CCHVAE1.001.001.001.001.00
DiCoFlex0.670.630.650.670.58
DICTUM0.740.870.800.870.81

DiCE and CCHVAE optimise each candidate per query until it is valid, so their pools are nearly all valid at a higher cost. DICTUM samples without per-query optimisation; still, every query has a valid candidate and all have at least k=10k=10, which is why the returned sets reach 1.00.

Thank you for the rebuttal, the additional experiments are meaningful, so I am wllling to raise my score.

Dear Reviewer, Thank you very much for your positive feedback and for acknowledging our additional experiments. We are extremely glad to hear that they successfully addressed your concerns. We noticed that the score in the system has not been updated yet. We would be very grateful if you could kindly reflect your mentioned score increase in the system whenever you have a moment.

Reviewer_QeQ6OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes DICTUM, a conditional diffusion model that generates diverse sets of counterfactual explanations for mixed-type tabular data. It trains on diverse opposite-class neighbors selected with a k-DPP, uses a Transformer-based diffusion backbone that models numerical and categorical features jointly, and enforces user-specified actionability constraints at inference with an SVDD-inspired branching strategy, without retraining. The experiments cover five tabular benchmarks and evaluate validity, proximity, sparsity, plausibility, and diversity, together with a constraint-strictness sweep and ablations.

Critical strengths
  1. The authors build on the latest diffusion models for tabular data and use them to generate counterfactuals that are both plausible and diverse. I think this is a good direction. Most counterfactual methods for tabular data either optimize one counterfactual or sample from a narrow distribution, and a strong generative backbone is what makes plausibility and diversity possible at the same time.

  2. I like the training setup. The model is conditioned on the query point, and the target is a distribution over likely valid counterfactuals (opposite-class neighbors) instead of a single optimal one. Because of this, user preferences and constraints can be handled at inference time without retraining, which is how recourse would actually be used in practice.

Critical weaknesses

1. For actionability, the evaluation does not show what the guidance really adds, and the metrics may hide changes to numerical features.

My concern here is mainly about actionability. Constraints are enforced in three places: the penalty γ·v in the branch score, the projection Πₓ, and the final feasibility filter A. For actionability constraints, there is a much simpler baseline that the paper does not compare against: replacement (inpainting), without branching. If a feature is not actionable, its value is known, so at each reverse step it can simply be overwritten, either with the original query value or with its noised version at the current step. The forward process is known, and the noise is already sampled at the start when the query is noised, so this is straightforward. Immutable categorical cells can simply be kept unmasked, and “increasing” features can be clamped to be at least their query value. Other types of constraints may really need guidance, but for actionability I can’t tell how much the SVDD-style guidance adds over this baseline, which is almost free.

Also, sparsity is computed only on categorical features and proximity only on continuous ones (L315–316, Table 5). A diffusion model will probably nudge every numerical feature a little, and neither metric would show it. I would like to see numerical sparsity with a tolerance (feature j counts as changed if |xj−x′jx_j − x′_j| > ε). Related to this is the SDEdit start at η = 0.6 (L273). In masked diffusion, a categorical cell that stays unmasked during forward noising never changes in the reverse process. So categorical sparsity may reflect η more than the model, and Fig. 6 (L755) seems consistent with that.

2. The method as described does not match what seems to have been run, and many hyperparameters are missing, which hurts clarity and reproducibility.

I noted the followig inconsistencies:

  • Branching procedure. §4 (L367–371) describes M = 5 full trajectories to t = 0, keeping the best (K = 1) and restarting for each counterfactual, with no k-DPP. Algorithm 1 and Table 6 (L681–700, L964) describe K = 5 branches per step, scored on the one-step prediction x̂₀, with a k-DPP at the end. The letters M and K also mean different things elsewhere (M = 100 pool, K = 10 outputs).

  • Candidate pool. L218 says the candidate pool consists of “all training examples from the target class y′,” but Table 6 (L940) says it is the 750 nearest neighbors by Gower distance (5 × 150). These are different procedures, and the text should describe the one that was actually used.

  • Post-hoc selection. Algorithm 1 (line 21) uses greedy k-DPP selection, but §3.4 (L269) says “k-DPP sampling” and §4 (L310) says “selection.” Greedy MAP is deterministic and sampling is not, so the text should say “greedy” everywhere.

I also found some important values missing:

  • σqσ_q and σsσ_s for the training-time k-DPP (L193). Table 6 gives them only for post-hoc selection.

  • σmin,σmax,εmin,εmax,σ_{min}, σ_{max}, ε_{min}, ε_{max}, and the initial ρjρ_j and kjk_j (L232–237).

  • The Gumbel-softmax temperature.

  • λnumλ_{num} and λcatλ_{cat} (L257).

Several fixed design choices are not ablated: k = 150 targets per query, the 20,000-sample training subset, the training-time kernel scales, and the branching settings λ and K. The ablations over T and η (Figs. 5–6) are appreciated.

Questions for authors

Clarifications that affect my assessment

1 In Eq. 7, α′tα′_t is never defined. I assume it is negative. The term log$⟨μ_θ^{cat}, x₀^{cat}′⟩$ is minimal when both vectors are orthogonal, so the sign of the weight matters. If α′t<0α′_t < 0, minimizing the loss pushes μθcatμ_θ^{cat} toward x0x₀. If the weight were positive, minimizing it would push μθcatμ_θ^{cat} to be orthogonal to x0x₀ instead. Please define α′tα′_t explicitly. With per-column schedules, please also write the sum over columns j.

  1. Which branching procedure produced the main results: the one described in §4 or Algorithm 1? And which candidate pool was used: all target-class training examples (L218) or the 750 nearest neighbors (L940)?

  2. The exact k-DPP is justified as too costly because it “must be repeated for every training instance in each epoch” (L198–200). But greedy MAP is deterministic, so the targets could be computed once and stored as a lookup table. With N = 750, exact sampling also seems affordable. Were the targets computed once or per epoch? Did you try exact sampling per epoch?

  3. Guidance scale (L666): is t continuous in [0,1] or the integer step? With integer steps and t0t₀ = ⌊0.6·49⌋ = 29, λtλ_t goes from about 584 down to 24. That makes the branch softmax basically an argmax at every step. Also, why is guidance strongest at the noisiest steps, where x^0x̂₀ is least reliable? A sweep over λ and K would help.

  4. Please define the projection Πₓ for each constraint type (L667–668). Does it paste in the clean xjx_j or xjx_j noised to the current step? How are categorical cells handled?

  5. Condition dropout of 0.25 (Table 6) suggests classifier-free guidance training, but CFG is never mentioned. Is CFG used at sampling? If yes, with what scale, and which conditions are dropped? Also, since the labels come from the classifier, y′ is fully determined by x, so what does conditioning on y′ add?

  6. Encoding (L299): categories are described as ordinal-encoded, but the diversity metric and Table 5 use one-hot. Which encoding does each component use (diffusion model, classifier, baselines, metrics)? For the “increasing” categorical constraints (Education, Pay_0, Pay_2), how is the category order defined?

Points on formulation

  1. Eq. 4 writes the reverse transition as a product of a numerical and a categorical factor, but pθnump_θ^{num} is never written out (L247). The text only says that μθnumμ_θ^{num} predicts the noise. The two network outputs also live in different spaces: μθnumμ_θ^{num} is a noise estimate in Rdnumℝ^{d_num}, while μθcat′μ_θ^{cat′} is a distribution on the probability simplex. In score-based form the product becomes a sum, and with the shared μθμ_θ notation this reads as adding network outputs that are not in the same space. I understand the formulation comes from prior work and that this is mostly notation. Still, I would present it more generally: define the joint reverse transition, then say explicitly how each factor is parameterized (for example, pθnump_θ^{num} as a Gaussian whose mean is derived from the predicted noise). Please also explain how this relates to DθD_θ in Algorithm 1, which predicts the clean x̂₀, and to the “stochastic, second-order correction” sampler in Table 6.

  2. Eq. 5 gives the reverse transition only for cells that are still masked. What happens to cells that are already unmasked: do they stay unchanged or can they change?.

Suggestion for future work (does not affect my assessment)

The idea of conditioning on the query point and learning a distribution over valid counterfactuals seems quite general. It would be interesting if the authors discussed whether it could transfer to other modalities, such as images or time series.

6 comments

Response to Reviewer QeQ6

We thank the reviewer for the detailed reading.

W1.1 + Q5. Inpainting baseline

We implemented it as described. In Setup 3, both it and SVDD + projection reach 100% constraint validity (unguided: λ=0\lambda=0 in W2.5). Selected sets differ by ≤0.01; inpainting is ~7× cheaper and we add it as a baseline. Branching matters for constraints that cannot be written into the state: adding "change at most 2 features" to Setup 3 on Credit Default, inpainting can only reject violating samples (20% of the pool is valid), while branching with a stochastic numerical step reaches 39%, at the same proximity and LOF. Q5. After each step, Πx\Pi_{\mathbf{x}} sets immutable cells to the clean query value and increasing ones to max⁡(xj′,xj)\max(x_j',x_j) (category order of Q7.2); pasting the noised value gives the same results.

W1.2. Numerical ε-sparsity

We add ε-sparsity to Table 1 (see 5tWg, W5.1). At ε=0 DICTUM changes all numerical features; at ε≥1% of range it is as sparse as or sparser than the kNN targets.

W1.3 + F2. η and unmasked cells

Unmasked cells are copied and never change; η bounds the masked fraction. Among masked cells the model changes categories less often than random fill from the marginal (η=0.6: 0.29–0.48 vs 0.34–0.63). Raising η increases validity (Fig. 6).

W2.1 + W2.2 + W2.3 + Q2.1 + Q2.2 + Q3. Branching, pool, selection

We apologise for the inconsistency: all results (Q=2000Q=2000 queries) use Algorithm 1: Table 1 with B=1B=1 and no constraints, Tables 2/4 with B=5B=5 branches scored on x^0\hat{\mathbf{x}}_0, then projection, filtering and kk-DPP; the score's validity term is the classifier's target-class probability. We remove the §4 paragraph (unused variant); symbols: ktr=150k_{tr}=150 targets, P=100P=100 pool, B=5B=5 branches, k=10k=10 outputs. Training candidates are the 5⋅150=7505\cdot150=750 Gower-nearest target-class examples (L218 corrected). Both selections are deterministic greedy MAP ("sampling" replaced); training targets are computed once ("every epoch" was wrong). Exact per-epoch sampling (affordable at N=750N=750) was not tried; targets matter less than selection.

W2.4. Missing values

σq=σs=0.5\sigma_q=\sigma_s=0.5; σmin⁡=0.002\sigma_{\min}=0.002, σmax⁡=80\sigma_{\max}=80, ϵmin⁡=10−5\epsilon_{\min}=10^{-5}, ϵmax⁡=10−3\epsilon_{\max}=10^{-3}; learnable ρj\rho_j, kjk_j initialised to 3.3 and 1.0 (TabDiff); Gumbel temperature 0.01; λnum=λcat=1\lambda_{num}=\lambda_{cat}=1. All added to Table 6.

W2.5 + Q4.1 + Q4.3. Ablations, λ, B

Constraint-valid pool without projection (%; Setup 3, constant λt\lambda_t, λ\lambda at B=5B=5, BB at λ=20\lambda=20; Paper = Table 2):

λ=01520100B=12510Paper
Adult33454442373338424443
Default20372922202021222325
Lending6766666666

Gains grow with BB; λ=1\lambda=1–5 is best. With projection, every setting and schedule gives 100% constraint validity; class validity varies by ≤2.5 pp.

Q4.1. tt is the normalised time ti=i/(T−1)∈[0,1]t_i=i/(T-1)\in[0,1], not the integer step. Table 2 uses λt=λ(t+0.2)\lambda_t=\lambda(t+0.2): from t0=0.6t_0=0.6 it falls from 16 to 4, so the branch softmax stays soft. Guidance is strongest early because the first steps decide which change a trajectory makes; starting from the noised query, x^0\hat{\mathbf{x}}_0 is already informative.

Q1. Eq. 7

αj,t=1−ϵmin⁡−(1−ϵmin⁡−ϵmax⁡)tkj\alpha_{j,t}=1-\epsilon_{\min}-(1-\epsilon_{\min}-\epsilon_{\max})t^{k_j}, so αj,t′≤0\alpha_{j,t}'\le0; the weight αj,t′/(1−αj,t)\alpha_{j,t}'/(1-\alpha_{j,t}) and the log term are both non-positive, so each term is a weighted cross-entropy of the true category (masked-diffusion NELBO, Sahoo et al., 2024). We define α′\alpha' and write the sum over columns.

Q6.1 + Q6.2. Dropout and y′y'

Dropout zeroes the condition for 25% of samples as a regulariser; CFG is not used. y′y' matters: pool validity is 74–87% with y′y'=target, 44–76% with zeros and 6–10% with y′y'=query class.

Q7.1 + Q7.2. Encoding, category order

Categoricals are integer codes, one-hot encoded everywhere (plus a mask token in the diffusion model). "Increasing" constraints compare category ranks: numeric order for PAY_*; string columns (e.g. education) are sorted semantically.

F1. Eq. 4

L247 was wrong: μθnum\boldsymbol\mu_\theta^{\text{num}} predicts clean values, not noise; the numerical step is a deterministic ODE step, not a Gaussian: xs=xt+(σ(s)−σ(t))dt\mathbf x_s=\mathbf x_t+(\sigma(s)-\sigma(t))\mathbf d_t, dt=(xt−μθnum)/σ(t)\mathbf d_t=(\mathbf x_t-\boldsymbol\mu_\theta^{\text{num}})/\sigma(t). Both heads predict the clean sample for disjoint columns: Eq. 4 factorises, outputs are never added. DθD_\theta denotes both heads; x^0\hat{\mathbf{x}}_0 uses the argmax of μθcat\boldsymbol\mu_\theta^{\text{cat}}. In Table 6, "second-order correction" is Heun; "stochastic" was a mislabel (no churn noise is added).

FW. Future work

On images, pixel-space neighbours are far less meaningful than tabular ones. We leave other modalities for future work.

Reviewer_QeQ6OpenReview ↗

Partially unreadable

Dear authors, could you please double-check your markdown for your answers for "W2.1 + W2.2 + W2.3 + Q2.1 + Q2.2 + Q3. Branching, pool, selection", "Q1. Eq. 7" and "F1. Eq. 4"? I cannot read these as there seems to be some formatting error.

Thank you for pointing this out, and apologies for the inconvenience. The response rendered correctly in our local Markdown preview, but OpenReview’s Markdown interpreted some underscores in the LaTeX subscripts as italics, which broke the equations. We have fixed the formatting and updated the response

Reviewer_5tWgOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

DICTUM uses a tabular-data specific diffusion process to learn mapping instances from one (negative) class to the target. The goal is to generate diverse counterfactual sets, while allowing for actionability constraints using SVDD-style guidance during inference. Results show better performance in terms of validity, proximity, plausibility, and diversity, with a tradeoff for sparsity.

Critical strengths

S1) The paper is well-written and easy to understand. Metrics, datasets, competitors, and related work are explained adequately. S2) The inference-time branching idea is interesting, and servers its purpose (actionability constraints)

Critical weaknesses

W1) The novelty appears limited. This work is very similar to the main competitor DiCoFlex, essentially swapping out the normalizing flows for an existing tabular diffusion method (Tabdiff), and changing kNN selection to k-DPP. The paper does not evaluate very basic ablations, like how DiCoFlex with k-DPP selection performs, or DICTUM with standard kNN selection (the one used by DiCoFlex). This baseline is necessary to isolate whether the reported improvements arise from the diffusion model or simply from the change in the selection mechanism.

W2) Regarding inference-time guidance for actionability constraints, I think that a similar heuristic could be applied to DiCoFlex, since it also samples CFs in a similar manner. This would lead to a more fair comparison in Table 2. W3) It’s not clear if the method requires access to the model during inference. A table describing basic algorithmic properties of all relevant methods (access to model gradients, weights, access to probabilities or hard decisions during inference, etc.) would be useful to understand how the baselines compare to the method.

W4) Regarding the MAP approximation (3.2). It is unclear if this is a contribution by the authors. If not, a citation is missing. If so, this process should be explained better (maybe in the Appendix), more intuitively to follow, but also formally e.g. by defining an algorithm. Parts of the implementation are not explained adequately and citations are missing, e.g. for Cholesky factor updates and Schur complement of L relative to Y.

W5) Numerical sparsity (epsilon-sparsity), which is reported by DiCoFlex, is not reported by the authors. Additionally, is cosine distance for diversity the correct/relevant metric, especially for one-hot categorical features?

Questions for authors
  • How influential is the selection of the Gower distance for the distance metric?
  • How sensitive is the method with regard to the Gower Categorical Weight hyperparameter?-
4 comments

Response to Reviewer 5tWg

We thank the reviewer for the careful reading and concrete suggestions.

W1. Novelty relative to DiCoFlex / TabDiff

TabDiff is an unconditional generator, turning it into an explainer required changes that it does not contain: conditioning on (x,y′)(\mathbf{x},y') and starting from the noised query, so samples stay close to x\mathbf{x}; masked categorical diffusion in which unmasked cells are copied, so categorical sparsity holds by construction; and black-box constraint handling at inference, by branching on x^0\hat{\mathbf{x}}_0, without retraining. Relative to DiCoFlex, the gain comes from the generator itself: even when only DiCoFlex gets our post-processing, DICTUM has lower LOF (0.15–0.17 vs. 0.28–0.41) and higher diversity (W2), and it supports monotone constraints (W3)

W2. Generator vs. selection

We would like to respectfully note that part of this ablation is already in the paper: Fig. 2 (left) isolates the training targets and Fig. 4 the post-hoc selection (kk-DPP / kNN / random / farthest on DICTUM's pool). We added DiCoFlex with our post-processing and DICTUM without it.

Table R1.

MethodProx. ↓Spars. ↓LOF ↓Diversity ↑
Adult
DiCoFlex0.890.290.450.35
DiCoFlex + post-proc.0.380.320.280.45
DICTUM − post-proc.0.340.290.170.64
DICTUM0.230.260.170.74
Credit Default
DiCoFlex0.600.250.300.34
DiCoFlex + post-proc.0.420.250.410.43
DICTUM − post-proc.0.350.230.150.63
DICTUM0.270.190.170.75

Post-processing improves DiCoFlex's proximity and diversity, but DICTUM remains better in proximity, LOF and diversity on both datasets, even without post-processing.

W3. Model access

DICTUM uses only predictions of hh, never gradients or weights: labels to relabel training data and filter samples, and the target-class probability to score branches under constraints.

DiCEC-CHVAEDiCoFlexDICTUM
Gradients / weights of hhno*nonono
hh output at inferenceprobabilitylabellabellabel (+ prob. with constraints)
Optimisation per queryyesyesnono
Immutable / monotone constraintsyes / yesno / noyes / noyes / yes

*Model-agnostic random-sampling DiCE.

W4. MAP approximation and Cholesky / Schur details

Both selections use the fast greedy MAP algorithm of Chen et al. (2018), based on the Schur-complement marginal gain (Kulesza & Taskar, 2012); it is not our contribution, and we add both citations.

W5. ε-sparsity

We calculated ε-sparsity (fraction of numerical features with ∣xj′−xj∣>0.05⋅rangej|x'_j-x_j|>0.05\cdot\mathrm{range}_j):

ε-Spars. ↓AdultBankDefaultGMCLending
DiCE0.160.130.060.100.15
CCHVAE0.140.230.160.190.55
DiCoFlex0.480.380.250.200.58
DICTUM0.230.360.210.260.39

Diffusion nudges several numerical features: DICTUM is less ε-sparse than DiCE/CCHVAE but sparser than DiCoFlex on 4/5 datasets, and kk-DPP selection lowers it by 0.07–0.14 vs. random. DiCE's sparsest sets come with the highest LOF (0.26–1.15 vs. 0.07–0.20).

Q1. Cosine diversity with one-hot categoricals

Cosine diversity measures directions of change independent of magnitude; magnitude-based diversity rewards sets that are simply far from the query (DiCE, whose proximity is 3–10× worse). We agree on its limitations: one-hot changes have norm 2\sqrt2, random sparse edits score high, and values are not comparable across datasets.

Q2. Gower distance and its categorical weight

Distance for training targets. We retrained DICTUM with alternative distances, including the MAD-L1 of Wachter et al. (2017)/DiCE and HVDM (Wilson & Martinez, 1997).

Table R2. Adult, mean over 3 seeds, validity 1.00.

DistanceProx. ↓Spars. ↓LOF ↓Diversity ↑
Gower (paper)0.210.250.160.72
L1 [num | one-hot]0.270.250.160.69
L1 (numerical only)0.180.280.150.68
MAD-L1 + mismatch0.200.260.160.71
HVDM0.350.240.180.67

Gower and MAD-L1 give the best balance; only HVDM is clearly worse.

Categorical weight, post-hoc selection, 0.75 (paper) → 0.25:

AdultBankDefaultGMCLending
Spars. ↓0.26 → 0.330.17 → 0.220.19 → 0.240.14 → 0.230.25 → 0.34
Diversity ↑0.74 → 0.700.72 → 0.710.75 → 0.740.92 → 0.900.75 → 0.74

Categorical weight, training targets (Adult / Default, seed 42):

WeightProx. ↓Spars. ↓LOF ↓Diversity ↑
0.00.19 / 0.240.30 / 0.210.15 / 0.160.70 / 0.75
0.75 (paper)0.23 / 0.270.26 / 0.190.17 / 0.170.74 / 0.75
1.00.24 / 0.270.26 / 0.190.16 / 0.170.74 / 0.75

In selection, validity, proximity and LOF change by ≤0.03 and the weight trades categorical sparsity for diversity, in training, it trades categorical sparsity for proximity. Both effects are smooth, with no sharp optimum, and the selection weight can be tuned at inference without retraining.

Dear Reviewer 5tWg, As the discussion period ends tomorrow, we would be grateful to know whether our response addresses your concerns. We would like to clarify one point, since our W2 heading covered the ablation rather than your question on applying the guidance to DiCoFlex. A normalizing flow has no intermediate states, so constraints that cannot be written directly into the sample can only be enforced by rejecting finished samples. With an added feature-change budget (Credit Default, Setup 3), rejection leaves 20% of the pool valid, compared with 39% with our branching (see our response to Reviewer QeQ6, W1.1). We are happy to clarify anything that remains.

Reviewer_5tWgOpenReview ↗

Thank you for the detailed response and additional experiments

Thank you for the detailed response and additional experiments. Some concerns are now resolved: the model-access table clarifies how DICTUM uses the classifier (W3), the greedy MAP inference is properly cited (W4), and the distance and Gower-weight analyses answer my questions. My main concerns remain.

W1. Where do the gains come from? Figure 2 shows that, within DICTUM, k-DPP training targets already increase diversity compared with kNN targets. However, DiCoFlex is still trained on kNN targets and receives only the post-hoc k-DPP selection. Likewise, DICTUM without post-processing is still trained on k-DPP targets. Thus, the models still differ in both the generator and their training targets.

A matched comparison is needed, ideally a 2×2 ablation over generator type (flow or diffusion) and training targets (kNN or k-DPP), with identical inference-time selection. Without this, the remaining diversity gain cannot be attributed to the diffusion generator. The lower LOF of DICTUM does suggest a benefit from diffusion, and I acknowledge that.

W2. Constraint handling. I agree that a flow has no intermediate diffusion states and cannot use the same branching procedure. However, the constraints evaluated in the paper can also be enforced on completed samples by resetting immutable features, clamping increasing features, and applying the same feasibility filter.

The response to Reviewer QeQ6 also shows that, in Setup 3, simple inpainting and branching produce results differing by at most 0.01, while inpainting is approximately seven times cheaper. Therefore, for the constraints actually evaluated, branching does not appear to explain the advantage. The additional feature-budget experiment shows that branching may help for other constraints, but that constraint is not part of the main comparison.

It remains unclear whether DiCoFlex received the same projection and filtering. If not, its poor validity and proximity in Setups 2 and 3 may reflect asymmetric constraint handling rather than an inherent limitation of the flow.

DiCE baseline. The access table in the response marks DiCE as the model-agnostic random-sampling variant. If this is the variant used in the experiments, it does not match the description in Section 4, where DiCE is “a per-instance optimization method that incorporates a determinantal point process diversity term in the loss”, since the random-sampling variant does not optimize this loss. It is a reasonable choice for a strict black-box comparison, but it should be named as such. Since the classifiers here are differentiable, general state-of-the-art claims would also need the optimization-based DiCE, with its additional access assumptions stated explicitly.

W5. Numerical sparsity and coverage. I appreciate the new ε-sparsity table. It shows that DICTUM is more numerically sparse than DiCoFlex on four of five datasets. However, at a tolerance of 5% of the feature range, it is less sparse than DiCE on all five datasets and less sparse than C-CHVAE on four. At ε = 0, it changes every numerical feature (response to Reviewer QeQ6). Therefore, the unqualified claim that DICTUM is competitive in sparsity with the state of the art is not supported. A more accurate claim would distinguish categorical sparsity from numerical sparsity.

The claim of substantially broader coverage is also supported only indirectly through cosine diversity. The authors acknowledge the limitations of this metric, and no direct measure of distributional coverage is reported. The results support higher diversity under the chosen metric, but not necessarily broader coverage of the counterfactual space.

I appreciate the authors’ effort and the concerns they resolved. However, the source of the diversity gain remains confounded, the constrained comparison with DiCoFlex is not demonstrably like-for-like, the general sparsity claim is not supported by the reported numerical results, and the DiCE baseline does not appear to match its description in the paper. I therefore keep my score.

Thank you for the follow-up.

W1. Source of the gain. Three of the four cells already exist, all with the same selection (random choice among valid candidates):

Prox.↓ / Div.↑AdultBankDefaultGMCLending
Flow, kNN0.89/0.350.89/0.340.60/0.340.36/0.520.67/0.33
Diffusion, kNN0.29/0.600.48/0.570.32/0.560.35/0.680.48/0.59
Diffusion, k-DPP0.34/0.640.51/0.610.35/0.630.37/0.770.50/0.64

With the same targets and selection, diffusion raises diversity by 0.16–0.26, more than the whole effect of the targets (0.04–0.09). It does so with margin loss of proximity. Diffusion with kNN targets and no k-DPP selection (0.60/0.56) also beats DiCoFlex with k-DPP selection (0.45/0.43). The missing flow + k-DPP cell cannot reverse this. We will add it.

W2. Constraints. DiCoFlex was run as described in its paper, and its finished samples went through the same projection and feasibility filter as ours, so DICTUM has no advantage in constraint handling. The difference is when the constraints act: inpainting and branching fix values in intermediate states, and the other features are then re-generated conditioned on them. Resetting a finished flow sample changes only the fixed features and leaves the rest unchanged, which breaks consistency. Branching also handles constraints that cannot be fixed in the sample.

DiCE. The random-sampling variant was chosen because all methods use only predictions. Gradient-based DiCE would have access no other method has. We will name it "DiCE (random sampling)", limit the claim to this setting, and report gradient-based DiCE separately.

W5. Section 4 defines Sparsity as categorical, so the claim matches the metric. We will label it "categorical sparsity". DiCE's numerical sparsity comes with 3–10× worse proximity and higher LOF (0.26–1.15 vs. 0.07–0.20). At ε ≥ 1%, DICTUM is as sparse as its real kNN neighbours, so changing many numerical features is typical of real data. We add ε-sparsity to Table 1. "Coverage" becomes "a broader set of recourse directions", supported by Figs. 2 and 4 in addition to cosine diversity.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.