acceptodds
Under review as a conference paper at ICLR 2027

Personalized Generative Recourse via Conjugate Cost Priors and Inverted Sampling

40 views

Abstract

Algorithmic recourse aims to provide individuals with actionable changes that yield a favorable prediction while remaining in-distribution and minimizing the cost of change. Recent generative recourse methods balance these criteria more robustly than inference-time optimization, but they fix a single cost function at training time, whereas the cost of a change is user-dependent. We present **ACCORD**, a generative diffusion model conditioned on the user's cost function, so that a single model can serve recourse tailored to each individual's notion of what a change costs. Rather than tilting a fixed generative model toward the user's cost at inference time, which is expensive and degenerates when that cost differs from the training cost, ACCORD amortizes training over a flexible conjugate prior that we design. Our cost prior linearly combines random cost curves and expresses the weighted, asymmetric, and range-dependent costs typical of recourse applications. Starting from only unpaired positive and negative instances, we exploit the conjugacy of the prior to invert the order in which (cost, target) training pairs are sampled, avoiding a search over positives for every drawn cost. We analyze the risk of the trained model and show that inverting the sampling lowers risk on costs well supported by the positive instances. Empirically, ACCORD outperforms both search-based recourse methods and diffusion guidance techniques under diverse user-specific costs.

Primary area:
generative models
Keywords:
Personalized recourse, amortization, diffusion models, propabilistic methods, sampling techniques, inference-time guidance

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_8fW8OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes ACCORD, a diffusion model that generates recourse conditioned on an individual’s cost function. Its main contribution is a tractable cost prior combining Gamma-distributed feature weights with random preference windows or categorical sets. Conjugacy enables an inverted training sampler: sample a positive target first, then a compatible cost, avoiding repeated cost-dependent searches over positive examples.

Critical strengths
  • Significance of considered problem. Serving heterogeneous preferences with one trained model is a practically relevant objective.
  • The strongest contribution is the combination of expressive, tractable cost priors with cost-marginalized pairing and posterior cost sampling. Equations 3–7 and Proposition 1(A) provide a clear mechanism for amortizing training-data construction across user preferences.
  • The paper evaluates both a synthetic setting with an available target distribution and three mixed-type tabular benchmarks.
Critical weaknesses
  • It is not consistently the best method. The paper makes broad superiority claims, but FACE achieves a higher overall score on COMPAS and HELOC in Table 1.
  • The test preferences may favor its approach. The main experimental description says user preferences are sampled based on compatibility with available positive examples. Real users may have difficult preferences that do not fit the data. Testing independently chosen preferences would better establish practical usefulness.
  • Missing related work or comparisons: The closest personalized amortized method, iCARMA, is discussed but not experimentally compared. Its causal setting makes direct comparison nontrivial, but a controlled comparison on shared capabilities would strengthen the contribution.
  • The comparison budgets differ. ACCORD uses 50 generation steps, while BFS and DFS use 16 timesteps with additional search operations. This does not automatically make the comparison unfair, but quality should be compared at similar runtime or computational cost.
Questions for authors
  • How should the coverage claims change under the constant-cost counterexample? What additional assumptions connect low coverage to low ESS and meaningful target-estimation error?
  • What is the actual candidate set used in experiments?
  • How are evaluation preferences generated for each dataset? Please specify the distributions of both weights and windows, the positive data used for posterior sampling, and results under independent preferences.
  • What exact reward do guidance baselines use?
5 comments

Thank you very much for your valuable feedback. We have uploaded a revised version of the paper incorporating your feedback. All the line numbers reported below are as per the revised version of the paper.

Q1: Coverage Claim under constant cost

The reviewer is right to point out that low coverage does not lead to low ESS in the constant-cost case, as stated below Equation (15). For this statement to hold, we additionally assume a normalization of the cost so that minx+∈D+CU(x+,x)=0min_{x⁺∈D⁺} C_U(x⁺,x) = 0, i.e., coverage is read relative to the cheapest recourse available in D+D^+. The bounds in Equation (15) are correct since both ESS^\widehat{ESS} and Z^(x,U)⋅eCˉ(x,U)\widehat{Z}(x, U) · e^{\bar{C}(x, U)} are shift invariant.

Q2: Candidate set used in experiments

Across all experiments, the candidate set N(x)\mathcal{N}(x) in Algorithm 1 is the entire positive-class training split, i.e., N(x)=Dh+\mathcal{N}(x)=\mathcal{D}_h^+. We do not apply any truncation (e.g., Top-(v) selection) to construct the candidate set. We have added this detail below Algorithm 1 at Line 950. If the reviewer had a different candidate set in mind, we would be happy to clarify.

Q3.1: Details of the distributions and data used

The evaluation preferences for each dataset are generated with the same posterior sampling as in training, using the test split of the dataset [Lines 406–407].

Distribution of weights and windows:

  • The prior distribution of the weight of each feature jj is λj∼Gamma(αj,βj)\lambda_j \sim \mathrm{Gamma}(\alpha_j, \beta_j) [Eq. 9, Line 235].
  • The posterior distribution of λj\lambda_j is Gamma(αj, βj+cj(mj(x+;x,ωj))(x+;x))\mathrm{Gamma}\big(\alpha_j,\ \beta_j + c_j^{(m_j(x^+;x,\omega_j))}(x^+; x)\big) [Eq. 11].
  • Each continuous window is defined as [lj,rj][l_j, r_j] [Line 210]. We define rj=ωj1r_j = \omega_{j1} and lj=ωj1−ωj2l_j = \omega_{j1} - \omega_{j2} to ensure lj<rjl_j < r_j [Lines 239–240]. The prior distribution of ωj1\omega_{j1} is Exp(Ωj1)\mathrm{Exp}(\Omega_{j1}) and that of ωj2\omega_{j2} is Exp(Ωj2)\mathrm{Exp}(\Omega_{j2}) [Line 244].
  • The posterior distributions of both ωj1\omega_{j1} and ωj2\omega_{j2} are mixtures of exponential distributions; for the detailed expressions, please refer to Lines 263–265.

Positive data used for posterior sampling: The positive test instances, reweighted by M(x+,x)=EU∼P(CU)[e−CU(x+,x)]M(x^+, x) = \mathbb{E}_{U \sim P(C_U)}\big[e^{-C_U(x^+, x)}\big], where x+x^+ is a positive test instance and xx is the rejected (negative) test instance [Eq. 3, Lines 162–164].

Q4: Reward used in guidance baselines

Based on the prior work, all baselines use a model trained for Pθ(x+∣x)∝Ph+exp−λˉ⋅ϕ(x,x+)P_\theta(x^+\mid x) \propto P_h^+exp^{-\bar{\lambda}\cdot\phi(x,x^+)}, on which we apply the guidance reward exp−CU(x,x+)+λˉ⋅ϕ(x,x+)exp^{-C_U(x,x^+) + \bar{\lambda}\cdot\phi(x,x^+)}, the second term is used to offset the tilt of the baselines. This makes the final sampling distribution as Ph+exp−λˉ⋅ϕ(x,x+)∗exp−CU(x,x+)+λˉ⋅ϕ(x,x+)=Ph+exp−CU(x,x+)P_h^+exp^{-\bar{\lambda}\cdot\phi(x,x^+)}*exp^{-C_U(x,x^+) + \bar{\lambda}\cdot\phi(x,x^+)} = P_h^+exp^{-C_U(x,x+)}. We have explicitly added this in App. F.2.1 (Line 1296).

Q3.2: Evaluation under Independent Sampling

We report results under posterior-sampled costs not to favor our method, but to evaluate performance under a balanced regime of covered (achievable) and low-coverage costs. As discussed in Sec. 3.3.1 (Lines 314–316), costs sampled independently (forward) of the data have little to no coverage in Dh+D_h^+. For instance, only approximately 15%, 10%, and 0% of independently sampled costs are covered on COMPAS, ADULT, and HELOC, respectively. In contrast, posterior sampling under inverted sampling provides a more balanced representation of both regions (Tab. 3). Thus, evaluating on posterior-sampled costs is a sensible choice, as it includes costs for which recourse is achievable on the data manifold.

DataSampling% of Covered costs sampled
ADULTFOW9.91%
INV54.95%
COMPFOW15.25%
INV85.12%
HELOCFOW0.00%
INV62.41%

Table 3: Percentage of covered costs sampled under Forward and Inverted Sampling. On high-dimensional HELOC, Forward Sampling predominantly generates unrealizable costs.

However, following the reviewer's suggestion, we repeat the evaluation under prior-sampled (Forward) preferences, where all methods struggle to satisfy user preferences due to low coverage (Table 11). Among generative methods, ACCORD and FOW satisfy more preferences than guidance-based baselines at lower cost, note that FOW is ACCORD without inverted sampling. FACE and DiCE achieve lower costs but suffer from substantially poorer validity, falling below random on two of the three datasets (both on ADULT and DiCE on COMPAS), making them less effective for recourse. This setting is largely low-coverage, as evidenced by low PRF and only 0–15% covered cost samples under Forward sampling (Table 3). ACCORD is also computationally more efficient to train than FOW, reducing complexity from O(∣Dh−∣⋅∣Dh+∣⋅n)O(|D_h^-|\cdot|D_h^+|\cdot n) to O(∣Dh−∣⋅(∣Dh+∣+n))O(|D_h^-|\cdot(|D_h^+|+n)) (Line 171-173).

        ADULT        COMPAS        HELOC
MethodPRF ↑InDist+ ↑Val ↑Cost ↓PRF ↑InDist+ ↑Val ↑Cost ↓PRF ↑InDist+ ↑Val ↑Cost ↓
RERD0.020.810.9124.100.110.740.973.520.001.000.9533.68
BFS0.040.800.9124.330.200.740.973.440.001.000.9534.31
DFS0.010.790.9124.980.140.700.973.660.001.000.9534.07
SVDD0.020.800.9124.410.110.690.963.580.001.000.9534.14
FACE0.100.740.2815.240.460.850.821.440.000.990.6420.53
DiCE0.110.700.3215.850.400.710.452.030.000.930.5122.65
FOW0.080.760.9017.170.270.750.982.320.000.990.9430.97
ACCORD0.090.760.8917.350.230.750.992.740.001.000.9630.45

Table 11 Comparing ACCORD against baselines across datasets under independently sampled preferences. FOW is ACCORD without inverted sampling.

Further, on ADULT, we use Mψ\mathcal{M}_\psi to partition preferences into covered and low-coverage regions and evaluate all methods separately in each (Table 12). ACCORD achieves one of the best cost–validity trade-offs in both regions (Fig. 3), whereas methods such as FACE and DiCE achieve lower costs at the expense of substantially poorer validity. These results demonstrate ACCORD's practical usefulness even under independent sampling and in low-coverage settings. We have incorporated these results into the paper (Fig. 3, Sec. 4.5, App. G.6).

        CoveredRegion        Low-coverageRegion
MethodPRF ↑Val ↑InDist+ ↑Cost ↓PRF ↑Val ↑InDist+ ↑Cost ↓
RERD0.190.920.8414.350.010.910.8124.31
BFS0.370.930.8411.930.030.910.8024.63
DFS0.200.910.8716.030.010.910.7925.40
SVDD0.200.960.7713.600.020.910.8024.76
FACE0.420.420.769.140.090.280.7415.44
DiCE0.490.420.767.350.100.320.6916.14
FOW0.520.940.787.880.060.900.7517.58
ACCORD0.560.920.786.350.070.890.7517.85

Table 12: ADULT under independently sampled preferences.

AuthorsOpenReview ↗

Addressing the weaknesses

FACE vs ACCORD

We acknowledge that FACE achieves a higher RScore on COMPAS and HELOC. However, RScore is a combined metric, that possibly does not give as much importance as it should to validity. Table 1 shows that FACE's validity is poor (0.85 and 0.65 vs. 0.99 and 0.96), making it ineffective for recourse. Its higher RScore is partly driven by lower recourse cost. But low cost without validity defeats the very purpose of recourse. A second limitation of FACE, is that it assumes access to the training data and searches within it during inference. This results in significantly higher computate overheads. We show the comparison below. FACE needs 25.17 and 57.01 ms/query on COMPAS and HELOC, vs. 1.24 and 0.57 for ACCORD (Fig 7: RScore vs Time). That makes it roughly 20× and 100× slower, since it searches at inference time. Worse, as we added a new RQ5 covering this comparison (App. G.7, Line 1550).

MethodADULT (ms/query)COMPAS (ms/query)HELOC (ms/query)
RERD17.0014.7823.20
BFS4.497.416.43
DFS1.405.922.10
SVDD3.905.025.54
DiCE523.03696.41132.84
FACE2237.0825.1757.01
ACCORD0.411.240.57

Table: Inference time comparison of ACCORD against baselines.

Comparison with iCARMA

We thank the reviewer for this suggestion. We have added a controlled comparison with iCARMA (App. G.1.1, Line 1330).

Setup. We compare against ACCORD Version 1, whose preference-weighted ℓ1\ell_1 cost corresponds directly to iCARMA's individualized action cost. The preferred-range cost of the later version of accord has no straightforward counterpart in iCARMA. Both methods share the same preference weights, classifier, train/test individuals, and parameter budget. As also mentioned by the reviewer, iCARMA needs a causal graph, since our datasets lack a it, we built a plausible causal ordering for each and derived an SCM from it. iCARMA does not report its validity-term weight λhinge\lambda_{hinge}, so we tested λhinge∈1,5,20,50\lambda_{hinge} \in \\{1, 5, 20, 50\\}. We show λhinge=20\lambda_{hinge}=20, which has the highest average validity; full results are in App. G.1.1.

        ADULT        COMPAS        HELOC
MethodInDist$^+$ ↑Val ↑Cost ↓InDist$^+$ ↑Val ↑Cost ↓InDist$^+$ ↑Val ↑Cost ↓
iCARMA (λhinge=20\lambda_{hinge}=20)0.350.592.120.710.460.620.870.531.05
ACCORD-v10.800.893.230.740.990.951.000.962.39

Findings. ACCORD achieves substantially higher validity on every dataset and higher plausibility (InDist$^+$), most notably on Adult. iCARMA finds cheaper interventions (lower cost), but many fail to change the outcome (only 0.45-0.60 validity). Tuning does not close the gap: raising λhinge\lambda_{hinge} improves iCARMA's validity at the cost of higher action cost and lower plausibility. This follows from its objective, where plausibility is only a soft regularizer. The iCARMA authors report the same validity–plausibility trade-off. ACCORD instead samples recourse from the positive data distribution, keeping validity and plausibility high.

Comparison budget discrepancy

To compare at a matched budget, we re-ran all guidance baselines (RERD, BFS, DFS, SVDD) with 50 sampling steps, the same as ACCORD. At this matched budget, ACCORD has the highest RScore and PRF and the lowest cost on all three datasets.

        ADULT        COMPAS        HELOC
MethodRScore ↑PRF ↑Val ↑Cost ↓RScore ↑PRF ↑Val ↑Cost ↓RScore ↑PRF ↑Val ↑Cost ↓
RERD0.910.180.9116.421.330.610.980.881.030.080.959.21
BFS1.100.390.9013.161.540.860.970.291.140.200.968.18
DFS0.950.230.9016.621.490.790.980.511.060.120.958.82
SVDD0.990.270.9114.641.280.590.910.851.050.100.968.93
ACCORD1.270.570.907.011.610.910.990.171.240.300.964.81

Table 14: ACCORD against the guidance baselines with every method run at 50 sampling steps.

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

This paper proposes ACCORD, a personalized generative recourse framework that conditions a diffusion model on user-specific cost preferences. The authors introduce a flexible cost representation using conjugate priors to capture weighted, asymmetric, and range-dependent preferences. They further develop an inverted triplet sampling strategy that efficiently constructs training data while prioritizing cost preferences supported by the positive data distribution. The paper provides theoretical analysis of the sampling strategy and evaluates the framework on synthetic and real-world datasets, demonstrating improvements in recourse quality, training efficiency, and preference adherence.

Critical strengths
  1. The paper addresses a clear and practically important limitation of existing generative recourse methods, which typically assume a fixed cost function during training. 2. The methodological contributions are well-integrated. The flexible cost representation, conjugate prior formulation, and inverted sampling strategy are designed to work together rather than being independent additions. 3. The inverted sampling strategy is a meaningful algorithmic contribution that reduces the computational cost of training triplet construction while preserving the conditional target distribution. 4. The theoretical analysis provides useful justification for the proposed sampling strategy, particularly regarding cost coverage and finite-sample risk. 5. The experimental evaluation is comprehensive and well-aligned with the methodological contributions. The authors investigate recourse quality, training efficiency, cost coverage, and robustness to prior specifications. 6. The results are generally convincing, showing improved recourse quality and faster training across multiple benchmarks, while maintaining competitive performance on other metrics.
Critical weaknesses
  1. The evaluation samples user preferences from a posterior distribution that favors well-supported costs. Additional evaluation using independently sampled or challenging user preferences would strengthen the generalization claims. 2. The main quantitative comparisons provide limited information about statistical variability, making it difficult to assess the significance of smaller performance differences.
Questions for authors
  1. How does ACCORD perform when test-time user preferences are sampled independently from the original cost prior, particularly for preferences with limited coverage in the positive dataset? 2. Could the authors provide repeated-run results with standard deviations or confidence intervals for the main recourse quality metrics to establish the reliability of the reported improvements?
3 comments
AuthorsOpenReview ↗

Rebuttal Response (1/2)

We thank the reviewer for their valuable feedback. We have uploaded a revised version of the paper incorporating your feedback. All the line numbers reported below are as per the revised version of the paper.

Q1: Evaluation under Independent Sampling

As suggested, we repeat the evaluation under prior-sampled (Forward) preferences, where all methods struggle to satisfy user preferences due to low coverage (Table 11). Among generative methods, ACCORD and FOW satisfy more preferences than guidance-based baselines at lower cost, note that FOW is ACCORD without inverted sampling. FACE and DiCE achieve lower costs but suffer from substantially poorer validity, falling below random on two of the three datasets (both on ADULT and DiCE on COMPAS), making them less effective for recourse. This setting is largely low-coverage, as evidenced by low PRF and only 0–15% covered cost samples under Forward sampling (refer Table 3). Furthermore, the reviewer notes that costs sampled from the posterior have good coverage. Table 3 in the paper reports the percentage of covered costs under both forward and inverted sampling, and shows that even under posterior sampling a sizeable fraction of costs is not covered. This shows posterior-based sampling provides a balanced amount of covered and low-coverage costs. ACCORD is also computationally more efficient to train than FOW, reducing complexity from O(∣Dh−∣⋅∣Dh+∣⋅n)O(|D_h^-|\cdot|D_h^+|\cdot n) to O(∣Dh−∣⋅(∣Dh+∣+n))O(|D_h^-|\cdot(|D_h^+|+n)) (Line 171-173).

        ADULT        COMPAS        HELOC
MethodPRF ↑InDist+ ↑Val ↑Cost ↓PRF ↑InDist+ ↑Val ↑Cost ↓PRF ↑InDist+ ↑Val ↑Cost ↓
RERD0.020.810.9124.100.110.740.973.520.001.000.9533.68
BFS0.040.800.9124.330.200.740.973.440.001.000.9534.31
DFS0.010.790.9124.980.140.700.973.660.001.000.9534.07
SVDD0.020.800.9124.410.110.690.963.580.001.000.9534.14
FACE0.100.740.2815.240.460.850.821.440.000.990.6420.53
DiCE0.110.700.3215.850.400.710.452.030.000.930.5122.65
FOW0.080.760.9017.170.270.750.982.320.000.990.9430.97
ACCORD0.090.760.8917.350.230.750.992.740.001.000.9630.45

Table 11 Comparing ACCORD against baselines across datasets under independently sampled preferences. FOW is ACCORD without inverted sampling.

Further, on ADULT, we use Mψ\mathcal{M}_\psi to partition preferences into covered and low-coverage regions and evaluate all methods separately in each (Table 12). ACCORD achieves one of the best cost–validity trade-offs in both regions (Fig. 3), whereas methods such as FACE and DiCE achieve lower costs at the expense of substantially poorer validity. These results show that ACCORD performs well when test-time user preferences are sampled independently from the cost prior, including preferences with limited coverage in the positive dataset. We have incorporated these results into the paper (Fig. 3, Sec. 4.5, App. G.6).

        CoveredRegion        Low-coverageRegion
MethodPRF ↑Val ↑InDist+ ↑Cost ↓PRF ↑Val ↑InDist+ ↑Cost ↓
RERD0.190.920.8414.350.010.910.8124.31
BFS0.370.930.8411.930.030.910.8024.63
DFS0.200.910.8716.030.010.910.7925.40
SVDD0.200.960.7713.600.020.910.8024.76
FACE0.420.420.769.140.090.280.7415.44
DiCE0.490.420.767.350.100.320.6916.14
FOW0.520.940.787.880.060.900.7517.58
ACCORD0.560.920.786.350.070.890.7517.85

Table 12: ADULT under independently sampled preferences. Left: covered region. Right: low-coverage region.

AuthorsOpenReview ↗

Rebuttal Response (2/2)

Q.2: Results Across Repeated Runs

We repeated every method with five different seeds and evaluated on the same test set. Table 14 reports the mean and standard deviation over the five runs. Across all methods and datasets the RScore standard deviations across runs are small (at most 0.06), and the RScore ranking of Table 1 (refer paper) remains unchanged. This suggests that the improvements reported in Table 1 are consistent across runs.

ADULTCOMPASHELOC
MethodRScore ↑PRF ↑Val ↑Cost ↓RScore ↑PRF ↑Val ↑Cost ↓RScore ↑PRF ↑Val ↑Cost ↓
RERD0.91 ± 0.000.18 ± 0.000.91 ± 0.0016.67 ± 0.061.33 ± 0.010.62 ± 0.000.98 ± 0.000.86 ± 0.011.03 ± 0.000.08 ± 0.000.95 ± 0.009.31 ± 0.05
BFS1.11 ± 0.000.39 ± 0.000.91 ± 0.0113.12 ± 0.241.58 ± 0.020.87 ± 0.010.98 ± 0.000.32 ± 0.021.14 ± 0.020.21 ± 0.010.95 ± 0.017.81 ± 0.06
DFS0.96 ± 0.010.25 ± 0.010.91 ± 0.0116.40 ± 0.181.48 ± 0.020.78 ± 0.010.98 ± 0.000.51 ± 0.011.08 ± 0.010.13 ± 0.010.95 ± 0.008.83 ± 0.08
SVDD0.93 ± 0.010.23 ± 0.010.92 ± 0.0015.24 ± 0.211.18 ± 0.060.50 ± 0.020.93 ± 0.031.01 ± 0.051.03 ± 0.010.09 ± 0.010.96 ± 0.008.65 ± 0.08
FACE0.73 ± 0.010.52 ± 0.000.28 ± 0.007.45 ± 0.061.72 ± 0.020.99 ± 0.000.84 ± 0.020.02 ± 0.011.46 ± 0.010.86 ± 0.010.65 ± 0.010.43 ± 0.02
DiCE0.76 ± 0.000.53 ± 0.000.35 ± 0.008.06 ± 0.031.11 ± 0.010.75 ± 0.010.49 ± 0.020.77 ± 0.151.09 ± 0.010.54 ± 0.000.58 ± 0.012.13 ± 0.04
FOW1.22 ± 0.010.53 ± 0.010.91 ± 0.007.34 ± 0.291.62 ± 0.010.92 ± 0.010.99 ± 0.000.16 ± 0.041.07 ± 0.010.14 ± 0.010.94 ± 0.016.80 ± 0.23
ACCORD1.28 ± 0.000.57 ± 0.010.90 ± 0.006.85 ± 0.111.61 ± 0.020.91 ± 0.000.99 ± 0.000.18 ± 0.011.23 ± 0.010.29 ± 0.010.96 ± 0.014.87 ± 0.04

Table 14: Comparing ACCORD against baselines across datasets: mean ± standard deviation over 5 runs on the same test set (± 0.00 denotes a standard deviation below 0.005).

Thank you for the detailed rebuttal and additional experiments. The evaluation under independently sampled preferences, including the covered and low-coverage analysis, adequately addresses my first concern. The repeated-run results with standard deviations also provide useful evidence of the consistency of the reported performance.

While some smaller performance differences remain inconclusive, I appreciate the authors' efforts in addressing my concerns and strengthening the experimental evaluation. I will maintain my original rating of 3

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

This paper proposes accord, a cost-conditioned diffusion model for personalized algorithmic recourse, allowing a single model to accommodate heterogeneous user preferences. It introduces a conjugate cost prior and an inverted sampling strategy to reduce training data construction costs and allocate more training samples to better-supported cost preferences.

Critical strengths

1.The paper addresses a meaningful problem: generating recourse tailored to individual cost preferences without retraining or inference-time search.

2.The conjugate cost formulation supports interpretable feature weights and range-dependent preferences, while enabling efficient sampling. 3. The paper provides theoretical analysis and evaluates the method on synthetic and real-world tabular datasets.

Critical weaknesses
  1. The theoretical guarantees may not hold for the actual implementation. The inverted-sampling analysis assumes sampling from the full positive dataset, whereas Algorithm 1 restricts candidates to a top-v subset selected using the prior-mean cost. This truncation generally changes the sampling distribution, and the paper does not establish that the claimed exactness and risk guarantees remain valid under this approximation.

  2. The risk analysis relies on a strong shared-optimum assumption. The claim that forward and inverted sampling recover the same optimum asymptotically depends on the existence of a common minimizer across queries. For a finite-capacity conditional diffusion model, different cost-conditioned targets may not be simultaneously representable. In this case, changing the sampling frequency can change the effective optimization trade-off, and the paper does not adequately characterize this effect.

  3. The real-world evaluation does not fully establish practical recourse quality. The experiments rely on synthetic labels from trained classifiers and a composite RScore, which may obscure trade-offs among validity, proximity, and preference adherence. Moreover, the method does not model causal dependencies between features, limiting the extent to which generated changes can be interpreted as actionable interventions in real-world settings.

Questions for authors
  1. Do the theoretical exactness and risk guarantees still hold when the implementation restricts positive candidates to a top-v subset?
  2. How does inverted sampling affect the optimal solution when a finite-capacity model cannot simultaneously fit all cost-conditioned targets?
  3. How can the authors validate that the generated recourse remains actionable when feature dependencies and causal effects are not modeled?
6 comments
AuthorsOpenReview ↗

Rebuttal Response (1/2)

We thank the reviewer for their valuable feedback. We have uploaded a revised version of the paper incorporating your feedback. All the line numbers reported below are as per the revised version of the paper.

Q1: Exactness and risk guarantees under top-vv truncation

In all reported experiments, candidates span the full Dh+(v=∣Dh+∣)D^+_h (v = |D^+_h|), so Proposition 11 applies to the evaluated implementation as stated; top-vv in Algorithm 11 is an optional scalability device, and we clarified this in the uploaded revision. (Please refer to Algorithm 11, Step 33.)

On the theory: Proposition 11(A) and (B) continue to hold exactly under truncation, with Dh+D^+_h replaced by the top-vv set N(x)\mathcal{N}(x) for both samplers, i.e., relative to the truncated plug-in target. That target differs from the full one by the tilted mass falling outside N(x)\mathcal{N}(x), which is largest for costs far from the prior mean, so truncation is an approximation. Proposition 11(C) will change quantitatively since the plug-in target is built from vv instead of the whole Dh+D^+_h and ESSESS is capped at vv, but it does not change the qualitative claim that coverage floors the achievable risk.

Finally, the training-time complexity reduction from O(∣Dh−∣⋅∣Dh+∣⋅n)O(|D_h^-|\cdot|D_h^+|\cdot n) to O(∣Dh−∣⋅(∣Dh+∣+n))O(|D_h^-|\cdot(|D_h^+|+n)) (Line 172) also assumes the full candidate set, consistent with the implementation used in all reported experiments.

Q2: The effect of Inverted sampling on optimal solution under finite model capacity

We agree that the assumption of a common minimizer is restrictive for a finite-capacity model, and in the revision we have added a statement in Proposition 11(A) (Line 297-300) that holds for any model class. Comparing the minimizers of the two losses, the excess loss of the inverted model relative to the forward one has non-positive covariance with relative coverage ρ(x,U)ρ(x,U) under the prior distribution, with equality iff each minimizer also minimizes the other loss, which includes the original common-minimizer setting as a special case. So changing the sampling frequency does change the optimum. The inverted model incurs lower loss on well-covered costs, and its excess loss concentrates on low-coverage costs, where Proposition 11(C) shows the risk floor is already high for any model. This matches the results we observe in Section 4.24.2 (Line 427-429), where inverted sampling improves recourse quality in the covered region on all datasets, and both methods provide lower-quality recourse in the low-coverage region.

Q3: Causality in the generated recourse

  • Unlike search-based methods, our model is supervised with real positive instances as recourse targets, so our generated examples are likely to preserve feature dependencies. We show the InDist+InDist^+ metric to validate this and our score ranges between 0.720.72 and 11 (measured using a LOF, which measures the conformity to the positive data distribution, refer Table 1313 under Section Appendix G.9G.9)

  • While the current tabular diffusion generator is not trained to preserve causal effects, our framework is compatible with causal generators. If a causal graph of the attributes is available, a causal generative model can be used to obtain causally valid recourse. To support this, we repeat our experiments on the semi-synthetic loan dataset of Karimi et al. [1]~[1], for which the true structural causal model (SCM) is available. We train an autoregressive generator via inverted sampling to generate the features in the causal order of this true SCM (refer to Section 4.64.6 and Appendix G.8G.8). We evaluate the generator by measuring the overlap between the true interventional distribution of the descendants of the actionable features restricted to the positive class and the generated conditional distribution of the descendants under the same intervention. We observe a substantial overlap between the generated distribution and the true interventional distribution. We report the plots in the revised version of the paper, refer Figure 44 and 88. The results show that when causal dependencies are modeled, a generator trained with Inverted sampling picks up the causal effects in the data.

  • Further, we would like to add that the main contributions of this work are:
    1. The design of a framework based on cost priors and cost-conditioned generators for handling inference-time costs
    2. Showing the benefits of Inverted sampling over Forward sampling (used in methods like iCARMA), both for efficient training (Sec 4.34.3, Line 453-456) and for better coverage of achievable costs (Table 22 and Table 33). We have additionally compared with iCARMA, on the suggestion of reviewer 8fW8, in Appendix G.1.1G.1.1.

 [1]~[1] Amir-Hossein Karimi, Julius von Kügelgen et al. "Algorithmic recourse under imperfect causal knowledge: a probabilistic approach." In Proceedings of the 34th International Conference on Neural Information Processing Systems, NIPS ’20.

AuthorsOpenReview ↗

Rebuttal Response (2/2)

Synthetic Labels from trained classifiers

We use labels from a trained classifier because the true labelling process (the oracle) is not available, so a generated recourse, being a new profile, cannot be checked against it. Following Garg et al. [2]~[2], we train a Random Forest on the real labels to act as the oracle, sample the experiment labels from it, and evaluate every generated recourse against it.

Composite RScore

Thank you for this feedback. We report RScore as a summary metric to avoid overloading the comparison across multiple datasets with too many metrics. However, Table 1 also reports validity (Val), proximity (Cost), and preference (PRF), keeping the trade-offs among these metrics visible. Table 13 (Appendix G.9) further reports all metrics separately along with standard deviations for every method.

 [2]~[2] Prateek Garg, Lokesh Nagalapatti, and Sunita Sarawagi. From search to sampling: Generative models for robust algorithmic recourse. In The Thirteenth International Conference on Learning Representations, 2025.

As the discussion phase is ending soon, we request you to please respond to our comments.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.