Pith. sign in

REVIEW 4 major objections 4 minor 43 references

Pareto Optimal Algorithmic Recourse in Multi-cost Function

T0 review · 4 major / 4 minor · reviewed 2026-08-08 · deepseek-v4-flash

Pith's one-line read A modified Bellman-Ford algorithm finds every Pareto optimal multi-cost recourse path in an actionability graph, with an epsilon-net procedure to scale the search to large graphs.

desk verdict A useful multi-cost recourse framing, but Algorithm 1 only returns exactly-η-hop paths, so Theorem 1 is false as stated; the scalability proof also has missing conditions. read the letter →

arxiv 2502.07214 v1 pith:2EFHFX3Y submitted 2025-02-11 cs.LG cs.AIcs.DS

classification cs.LGcs.AIcs.DS MSC 90C2990C3568R10
keywords algorithmicrecoursemulti-objectiveoptimizationParetooptimalityshortestpathepsilon-netnon-differentiablecostactionabilitygraphdynamicprogramming
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper proposes that algorithmic recourse—the problem of finding minimal-cost actions that flip a model's decision—should be solved as a multi-objective shortest-path problem on an actionability graph whose edges carry several cost values at once. The central claim is that a modified Bellman-Ford dynamic program, using Pareto tables that prune dominated paths, returns every Pareto optimal recourse path from a given starting point to any positively classified endpoint, for non-differentiable and discrete metric cost functions. Because the exact algorithm can be slow on large graphs, the paper adds an epsilon-net based graph-shrinking procedure that randomly samples a small representative set while preserving an approximation guarantee. A sympathetic reader would care because current recourse methods rely on gradient heuristics with a single blended cost, which can miss trade-offs between money, time, and effort.

What carries the argument

The carrying object is the Pareto table $D_v^\ell$: the set of nondominated cost vectors for paths from $s$ to $v$ that use exactly $\ell$ edges, updated by concatenating edge weights and pruning dominated vectors. This table lets the algorithm track trade-offs among $k$ cost functions without assuming differentiability. For scalability, the paper introduces $\kappa$-shrinkable vertices and uses $\epsilon$-nets over range spaces whose VC-dimension bounds the random sample size.

What would settle it

Take a graph with source $s$, target $t$, a direct edge $s\to t$ with cost $(1,1)$, and a two-hop path $s\to a\to t$ with cost $(2,2)$. Set $\eta=2$. The direct path is Pareto optimal but appears only in $D_1^t$, which Algorithm 1 discards when it returns $D_2^t$ at line 11; running the algorithm and observing that the reported set omits the direct path would contradict Theorem 1 as stated.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is Theorem 1: Algorithm 1 finds all Pareto optimal paths from source $s$ to any endpoint $t$ with $h(t)=1$, with time complexity $O(\gamma\tau\log^{k-4}\log\log\tau\cdot\eta|E|)$, where $\tau$ is the maximum Pareto table size, $\gamma$ the maximum degree, and $\eta$ the maximum number of allowed edges. The paper further claims that a random sample of size $O((|VC|^*/\epsilon)\log(1/\epsilon))$ is an $\epsilon$-net for the actionability graph, so running the algorithm on the sampled graph yields Pareto recourse paths that approximate the original ones within the shrinking factor $\kappa$.

Load-bearing premise

The theorem assumes that every Pareto optimal recourse path uses exactly $\eta$ edges, because Algorithm 1 returns only the table built at iteration $\eta$; the proof of correctness is deferred to a full version and is not supplied here.

Editorial extensions

If this is right

  • Users can define several metric cost functions, such as money, time, effort, and plausibility, and receive a set of alternative recourse plans showing trade-offs rather than a single blended optimum.
  • The method applies to models with discrete, non-smooth feature changes, such as credit-score bands or education levels, where gradient-based counterfactual methods are not viable.
  • The epsilon-net result implies that large real-world datasets can be randomly subsampled with a provable approximation factor, making the Pareto recourse search scalable.
  • If Theorem 1 holds, this is a polynomial-time exact algorithm for multi-cost recourse, giving an interpretable and theoretically grounded alternative to heuristic recourse generators.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The exactness claim as written applies only to paths using exactly $\eta$ edges; a natural extension is to take the union of tables across all path lengths so that shorter Pareto-optimal plans are not dropped, since Algorithm 1 returns only $D_\eta$.
  • The epsilon-net guarantee is demonstrated for costs that are within a constant factor of the data-space metric; for arbitrary metric costs without that structure, the sampling bound may not transfer, so practitioners should verify that their costs satisfy a Lipschitz-like condition.
  • Connecting to the multi-objective shortest path literature, the pruning step's complexity could likely be improved with output-sensitive algorithms, making the worst-case bound conservative in practice.
  • A testable extension would be to benchmark whether the random $\epsilon$-net preserves the full Pareto frontier on standard recourse datasets, rather than only a few individual path qualities.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes a multi-objective algorithmic recourse framework. It constructs an actionability graph whose edges carry multiple cost values, and Algorithm 1 (a Bellman-Ford-style dynamic program) is claimed to return all Pareto-optimal paths of at most eta hops from a negatively classified source to any positively classified endpoint. To handle large graphs, Section 4 introduces a notion of kappa-shrinkable vertices and proposes the use of epsilon-nets, claiming that a random sample of size O((|VC|^*/epsilon) log(1/epsilon)) yields a shrunk graph whose Pareto paths approximate the original ones. Experiments on MNIST and Adult datasets illustrate the trade-offs between criteria and the effect of sample size. The central theoretical claims are Theorem 1 (exactness of Algorithm 1) and the epsilon-net approximation guarantee.

Significance. If correct, the paper would address a real gap: most algorithmic recourse methods assume differentiable single cost functions, whereas recourse in practice involves multiple, possibly discrete, non-differentiable costs. The path-based formulation and the use of epsilon-nets for scalability are promising directions, and the paper explicitly frames falsifiable predictions (all Pareto paths found; approximation via random nets). The experiments, however, are only illustrative and do not compensate for the theoretical gaps. The main result, Theorem 1, is stated without proof and, as written, is false because the algorithm returns only exactly eta-hop paths. The scalability argument in Section 4 has a logically incorrect inference about kappa-shrinkability and the epsilon-net range space. Because these are load-bearing for the paper's contribution, the current version does not establish its claims.

major comments (4)
  1. [Section 3, Algorithm 1] Theorem 1 states that Algorithm 1 finds all Pareto-optimal paths from the source to any endpoint t with h(t)=1, while Section 3 defines eta as the maximum number of edges in a feasible path. The pseudocode, however, returns only D^eta_t (line 11), and D^ell_v is constructed solely by extending D^{ell-1}_u; there is no carry-over of D^{ell-1}_v into D^ell_v and no union over ell <= eta. Consequently, any nondominated recourse path using fewer than eta edges is missing. For example, with eta=2, edges s->t of cost (1,1) and s->a->t of cost (2,2), the algorithm returns D^2_t = {(2,2)} and omits the Pareto-optimal path (1,1). Since the proof of Theorem 1 is deferred to the 'full version,' the rendering shown cannot support the stated guarantee.
  2. [Section 4, Definition 1] Definition 1 only requires preservation of incoming edges: for all (p,i) in E, the shrunk vertex j must satisfy (p,j) in E and c(p,j) <= kappa c(p,i). It imposes no condition on outgoing edges (i,q). After shrinking i to j, any path that enters i and then leaves via (i,q) may have no counterpart in the shrunk graph unless (j,q) exists with a bounded cost. Thus the assertion that 'any G_S preserves the kappa-approximation factor' and the subsequent claim of a factor-kappa^l distance bound are not justified. Moreover, no argument is given that preserving distances implies preserving the Pareto-optimal path set under multiple cost criteria; the paper needs a formal statement relating Pareto nondomination in G and G_S.
  3. [Section 4, Demonstration] The demonstration claims that the condition Delta(i,j) <= c(i,j) <= kappa Delta(i,j) 'implies that the two points are kappa-shrinkable if and only if they are within the distance of 1 in the data space.' This is incorrect. The kappa-shrinkability condition involves the ratio c(p,j)/c(p,i); an upper bound on Delta(i,j) does not control this ratio when c(p,i) is small. Even if Delta(i,j) <= 1, one can have c(p,j)/c(p,i) arbitrarily large. Additionally, the existence of edge (p,j) is an actionability property, not a metric consequence. Therefore the claimed equivalence, and with it the VC-dimension bound based on balls in the data space, does not follow. The argument also conflates a multiplicative bound (kappa) with an additive one; the two notions require different treatment.
  4. [Section 3, Complexity analysis] The stated complexity O(gamma tau log^{k-4} log log gamma tau * eta |E|) is not well-defined for k=3, since log^{k-4} becomes log^{-1}, and it does not match the case analysis for k=2 and k=3 given in the same paragraph. More importantly, the analysis assumes the Pareto table size tau is a given parameter, but no bound is provided on how tau grows over eta iterations; in multiobjective shortest path, the number of nondominated labels can be exponential in the number of hops. Without such a bound, the claim that the algorithm is polynomial in the graph size is not established.
minor comments (4)
  1. [Section 4] The text contains a placeholder 'Section??' in the sentence referring to the scalability solution, and Section 3 defers the correctness proof to 'the full version of this work.' These should be resolved before submission.
  2. [Section 4, Definition 3] The definition 'shrunk by more than epsilon n vertices' is ambiguous: it is unclear whether the count refers to the number of original vertices that collapse into the representative, or the total number of vertices in the sample, and the phrasing should be made precise.
  3. [Section 5.3, Figure 4] The x-axis of Figure 4 is labeled 'cost1' but the text describes it as 'the criterion of cost1'; it would be helpful to specify the units (e.g., maximum number of digit transformations) and to report error bars or confidence intervals. The claim that the results 'support the argument of the epsilon-net' is not directly tested, since no comparison to the full graph is provided.
  4. [General] There are several typos, including 'probaility' in Section 4 and 'conclution' in Section 6; the paper would also benefit from a careful pass to ensure all references are formatted consistently.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity; the paper's claims rest on an algorithmic construction and standard external theorems, not on self-citation or fitted inputs.

full rationale

The paper does not fit parameters and then rename them as predictions; its experiments merely demonstrate trade-offs and sampling trends, which are not derived from the algorithm's assumptions by construction. Theorem 1 claims Algorithm 1 finds all Pareto optimal paths, but the proof is explicitly deferred to the full version; a missing proof is a correctness gap, not circularity. The scalability argument uses the standard Haussler-Welzl epsilon-net theorem as an external, machine-independent result, and the statement that any GS preserves a kappa-approximation factor follows directly from Definition 1 by the inequality c(p,j) <= kappa c(p,i) applied edge-by-edge along a mapped path. This is a definitional implication, not a fitted input called a prediction. There are no self-citations that carry the argument, no uniqueness theorem imported from the authors' prior work, and no known result renamed as unification. The strongest caveat is that Theorem 1 is unproved in the visible manuscript and Algorithm 1's described DP may fail to carry over shorter paths without a union step, but that concerns correctness, not circularity. Accordingly, the circularity score is 0.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The central claims rest on five unproved or domain-specific assumptions: metric costs with additive aggregation, complete edge enumeration, a deferred proof of Theorem 1, an incomplete shrinking rule, and a Lipschitz-like cost bound for the epsilon-net demonstration. The eta, kappa, epsilon, and delta hyperparameters are user-chosen and load-bearing. No physical entities are invented.

free parameters (4)
  • eta (max path hops) = user-chosen, 1 <= eta <= n-1
    Defines the maximum length of a recourse path; the algorithm returns only exactly eta-hop paths, so this choice changes the result set.
  • kappa (approximation factor) = user-chosen
    Defines kappa-shrinkable vertices and the claimed approximation ratio of the shrunk graph.
  • epsilon (net granularity) = user-chosen, 0 < epsilon < 1
    Controls the epsilon-net sample size in the Haussler-Welzl bound.
  • delta (failure probability) = treated as constant
    Appears in the sample size bound in Equation 1; the paper assumes delta is constant.
assumptions (6)
  • domain assumption User-supplied cost functions are metrics and path cost aggregation is additive.
    Section 2 permits aggregation 'such as summation or any maximum function', but Algorithm 1 concatenates edge cost vectors by addition; max aggregation is not implemented.
  • domain assumption All feasible edges between every pair of accessible points can be enumerated and stored.
    Section 3 constructs the actionability graph by iterating over each pair of vertices; this O(n^2) assumption is the main scalability bottleneck the paper then tries to remove.
  • ad hoc to paper Theorem 1 is true as stated in the full version.
    The manuscript writes 'the correctness of Algorithm 1 is provided in the full version of this work' and gives no proof, so the main exactness claim is an unverified assumption here.
  • ad hoc to paper Removing a kappa-shrinkable vertex preserves all relevant paths.
    Definition 1 only guarantees replacement edges for predecessors (p,i); the claimed approximation of shortest paths also requires replacement edges for successors (i,q), which is not established.
  • domain assumption Cost functions satisfy Delta <= c <= kappa Delta relative to the data-space metric.
    Section 4 uses this two-sided Lipschitz-like condition to argue that kappa-shrinkable points are exactly those within distance 1; this is not required by the 'free to choose metric costs' framing.
  • standard math Haussler-Welzl epsilon-net theorem applies to the shrinking range space.
    Section 4 invokes the theorem and estimates VC dimension as d+1 for balls; the application needs finite VC dimension of the range space for arbitrary metric costs, which is only demonstrated for a special ball case.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Pareto Optimal Algorithmic Recourse in Multi-cost Function." pith.science (2026). https://pith.science/paper/2EFHFX3Y

@misc{pith2026250207214,
  author       = {Pith},
  title        = {Pith review of: Pareto Optimal Algorithmic Recourse in Multi-cost Function},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2EFHFX3Y}},
  note         = {Machine review of arXiv:2502.07214}
}
read the original abstract

In decision-making systems, algorithmic recourse aims to identify minimal-cost actions to alter an individual features, thereby obtaining a desired outcome. This empowers individuals to understand, question, or alter decisions that negatively affect them. However, due to the variety and sensitivity of system environments and individual personalities, quantifying the cost of a single function is nearly impossible while considering multiple criteria situations. Most current recourse mechanisms use gradient-based methods that assume cost functions are differentiable, often not applicable in real-world scenarios, resulting in sub-optimal solutions that compromise various criteria. These solutions are typically intractable and lack rigorous theoretical foundations, raising concerns regarding interpretability, reliability, and transparency from the explainable AI (XAI) perspective. To address these issues, this work proposes an algorithmic recourse framework that handles non-differentiable and discrete multi-cost functions. By formulating recourse as a multi-objective optimization problem and assigning weights to different criteria based on their importance, our method identifies Pareto optimal recourse recommendations. To demonstrate scalability, we incorporate the concept of epsilon-net, proving the ability to find approximated Pareto optimal actions. Experiments show the trade-off between different criteria and the methods scalability in large graphs. Compared to current heuristic practices, our approach provides a stronger theoretical foundation and better aligns recourse suggestions with real-world requirements.

Figures

Figures reproduced from arXiv: 2502.07214 by the authors.

Figure 1
Figure 1. In this example, our approach provides three different paths which are all [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The paths in MNIST under different criteria of [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Multiple different paths from a starting data point to an end data point. [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: The trend among different sampling sizes shows that the quality of the [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 35 canonical work pages

  1. [1]

    Combinatorial and computational geometry52(1), 1–30 (2005)

    Agarwal, P.K., Har-Peled, S., Varadarajan, K.R., et al.: Geometric approximation via coresets. Combinatorial and computational geometry52(1), 1–30 (2005)

  2. [2]

    In: International Working Conference on Requirements Engineering: Foundation for Software Qual- ity

    Balasubramaniam, N., Kauppinen, M., Hiekkanen, K., Kujala, S.: Transparency and explainability of ai systems: ethical guidelines in practice. In: International Working Conference on Requirements Engineering: Foundation for Software Qual- ity. pp. 3–18. Springer (2022) Pareto Optimal Algorithmic Recourse in Multi-cost Function XIII

  3. [3]

    UCI Machine Learning Repository (1996), DOI: https://doi.org/10.24432/C5XW20

    Becker, B., Kohavi, R.: Adult. UCI Machine Learning Repository (1996), DOI: https://doi.org/10.24432/C5XW20

  4. [4]

    Quarterly of applied mathematics16(1), 87–90 (1958)

    Bellman, R.: On a routing problem. Quarterly of applied mathematics16(1), 87–90 (1958)

  5. [5]

    In: International conference on machine learning

    Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for con- trastive learning of visual representations. In: International conference on machine learning. pp. 1597–1607. PMLR (2020)

  6. [6]

    Genome medicine13, 1–16 (2021)

    Chereda, H., Bleckmann, A., Menck, K., Perera-Bel, J., Stegmaier, P., Auer, F., Kramer, F., Leha, A., Beißbarth, T.: Explaining decisions of graph convolutional neural networks: patient-specific molecular subnetworks responsible for metastasis prediction in breast cancer. Genome medicine13, 1–16 (2021)

  7. [7]

    In: International Conference on Parallel Problem Solving from Nature

    Dandl, S., Molnar, C., Binder, M., Bischl, B.: Multi-objective counterfactual ex- planations. In: International Conference on Parallel Problem Solving from Nature. pp. 448–469. Springer (2020)

  8. [8]

    arXiv preprint arXiv:2010.00711 (2020)

    Danilevsky, M., Qian, K., Aharonov, R., Katsis, Y., Kawas, B., Sen, P.: A sur- vey of the state of explainable ai for natural language processing. arXiv preprint arXiv:2010.00711 (2020)

Show all 43 references
  1. [9]

    Princeton University Press (2021)

    Daston, L., Galison, P.: Objectivity. Princeton University Press (2021)

  2. [10]

    Daws, R.: Medical chatbot using openai’s gpt-3 told a fake patient to kill themselves (2020), https://www.artificialintelligence-news.com/2020/10/ 28/medical-chatbot-openai-gpt3-patient-kill-themselves/

  3. [11]

    In: Proceedings of the 3rd innovations in theoretical computer science con- ference

    Dwork, C., Hardt, M., Pitassi, T., Reingold, O., Zemel, R.: Fairness through aware- ness. In: Proceedings of the 3rd innovations in theoretical computer science con- ference. pp. 214–226 (2012)

  4. [12]

    Journal of Machine Learning Research24(360), 1–37 (2023)

    Fokkema, H., de Heide, R., van Erven, T.: Attribution-based explanations that provide recourse cannot be robust. Journal of Machine Learning Research24(360), 1–37 (2023)

  5. [13]

    In: Proceedings of the sixteenth annual ACM symposium on Theory of computing

    Gabow, H.N., Bentley, J.L., Tarjan, R.E.: Scaling and related techniques for geom- etry problems. In: Proceedings of the sixteenth annual ACM symposium on Theory of computing. pp. 135–143 (1984)

  6. [14]

    IEEE Technology and Society Magazine41(2), 84–89 (2022)

    Gardner, A.: Responsibility, recourse, and redress: A focus on the three r’s of ai ethics. IEEE Technology and Society Magazine41(2), 84–89 (2022)

  7. [15]

    Hamilton, I.A.: An ai tool which reconstructed a pixelated pic- ture of barack obama to look like a white man perfectly illustrates racial bias in algorith (2020), https://www.businessinsider.com/ depixelator-turned-obama-white-illustrates-racial-bias-in-ai-2020-6

  8. [16]

    In: Proceedings of the second annual symposium on Computational geometry

    Haussler, D., Welzl, E.: Epsilon-nets and simplex range queries. In: Proceedings of the second annual symposium on Computational geometry. pp. 61–71 (1986)

  9. [17]

    Theoretical Com- puter Science 412(17), 1614–1628 (2011)

    Jiang, P., Chen, Y.p.: Free lunches on the discrete lipschitz class. Theoretical Com- puter Science 412(17), 1614–1628 (2011)

  10. [18]

    arXiv preprint arXiv:2010.04050 (2020)

    Karimi, A.H., Barthe, G., Schölkopf, B., Valera, I.: A survey of algorith- mic recourse: definitions, formulations, solutions, and prospects. arXiv preprint arXiv:2010.04050 (2020)

  11. [19]

    In: Proceedings of the 2021 ACM conference on fairness, accountability, and transparency

    Karimi, A.H., Schölkopf, B., Valera, I.: Algorithmic recourse: from counterfactual explanations to interventions. In: Proceedings of the 2021 ACM conference on fairness, accountability, and transparency. pp. 353–362 (2021)

  12. [20]

    Advances in neural information processing systems33, 265–277 (2020)

    Karimi, A.H., Von Kügelgen, J., Schölkopf, B., Valera, I.: Algorithmic recourse under imperfect causal knowledge: a probabilistic approach. Advances in neural information processing systems33, 265–277 (2020)

  13. [21]

    BIT Numerical Mathematics 28, 227–241 (1988) XIV W

    Karlsson, R.G., Overmars, M.H.: Scanline algorithms on a grid. BIT Numerical Mathematics 28, 227–241 (1988) XIV W. Chen et al

  14. [22]

    Frontiers in Physics10, 879171 (2022)

    Lawless, W.: Toward a physics of interdependence for autonomous human-machine systems: The case of the uber fatal accident, 2018. Frontiers in Physics10, 879171 (2022)

  15. [23]

    ATT Labs [Online]

    LeCun, Y., Cortes, C., Burges, C.: Mnist handwritten digit database. ATT Labs [Online]. Available: http://yann.lecun.com/exdb/mnist2 (2010)

  16. [24]

    In: Proceedings of the 2020 conference on fairness, accountability, and transparency

    Mothilal, R.K., Sharma, A., Tan, C.: Explaining machine learning classifiers through diverse counterfactual explanations. In: Proceedings of the 2020 conference on fairness, accountability, and transparency. pp. 607–617 (2020)

  17. [25]

    In:InternationalConferenceonArtificialIntelligenceandStatistics.pp.4679–4698

    Nguyen, D., Bui, N., Nguyen, V.A.: Feasible recourse plan via diverse interpolation. In:InternationalConferenceonArtificialIntelligenceandStatistics.pp.4679–4698. PMLR (2023)

  18. [26]

    IEEE transactions on neural networks and learning systems 32(2), 604–624 (2020)

    Otter, D.W., Medina, J.R., Kalita, J.K.: A survey of the usages of deep learning for natural language processing. IEEE transactions on neural networks and learning systems 32(2), 604–624 (2020)

  19. [27]

    IEEE Access 8, 220121–220139 (2020)

    Peres,R.S.,Jia,X.,Lee,J.,Sun,K.,Colombo,A.W.,Barata,J.:Industrialartificial intelligence in industry 4.0-systematic review, challenges and outlook. IEEE Access 8, 220121–220139 (2020)

  20. [28]

    ACM Computing Surveys (CSUR)51(5), 1–36 (2018)

    Pouyanfar, S., Sadiq, S., Yan, Y., Tian, H., Tao, Y., Reyes, M.P., Shyu, M.L., Chen, S.C., Iyengar, S.S.: A survey on deep learning: Algorithms, techniques, and applications. ACM Computing Surveys (CSUR)51(5), 1–36 (2018)

  21. [29]

    In: Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society

    Poyiadzi, R., Sokol, K., Santos-Rodriguez, R., De Bie, T., Flach, P.: Face: feasible and actionable counterfactual explanations. In: Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society. pp. 344–350 (2020)

  22. [30]

    Springer Science & Business Media (2012)

    Preparata, F.P., Shamos, M.I.: Computational geometry: an introduction. Springer Science & Business Media (2012)

  23. [31]

    In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H

    Rawal, K., Lakkaraju, H.: Beyond individualized recourse: Interpretable and inter- active summaries of actionable recourses. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) Advances in Neural Information Processing Systems. vol. 33, pp. 12187–12198. Cu...

  24. [32]

    In: Proceedings of the conference on fairness, accountability, and transparency

    Russell, C.: Efficient search for diverse coherent explanations. In: Proceedings of the conference on fairness, accountability, and transparency. pp. 20–28 (2019)

  25. [33]

    Advances in neural information processing systems34, 62–75 (2021)

    Slack, D., Hilgard, A., Lakkaraju, H., Singh, S.: Counterfactual explanations can be manipulated. Advances in neural information processing systems34, 62–75 (2021)

  26. [34]

    arXiv preprint arXiv:2206.03585 (2022)

    Srivastava, G., Jhaveri, R.H., Bhattacharya, S., Pandya, S., Maddikunta, P.K.R., Yenduri, G., Hall, J.G., Alazab, M., Gadekallu, T.R., et al.: Xai for cybersecu- rity: state of the art, challenges, open issues and future directions. arXiv preprint arXiv:2206.03585 (2022)

  27. [35]

    IEEE Transactions on Neural Networks and Learning Systems (2022)

    Tang, Y., Zhao, C., Wang, J., Zhang, C., Sun, Q., Zheng, W.X., Du, W., Qian, F., Kurths, J.: Perception and navigation in autonomous systems in the era of learning: A survey. IEEE Transactions on Neural Networks and Learning Systems (2022)

  28. [36]

    IEEE Access10, 14076–14119 (2022)

    Turay, T., Vladimirova, T.: Toward performing image classification and object detection with convolutional neural networks in autonomous driving systems: A survey. IEEE Access10, 14076–14119 (2022)

  29. [37]

    Advances in Neural Information Processing Systems 34, 16926–16937 (2021)

    Upadhyay, S., Joshi, S., Lakkaraju, H.: Towards robust and reliable algorithmic recourse. Advances in Neural Information Processing Systems 34, 16926–16937 (2021)

  30. [38]

    In: Proceedings of the conference on fairness, accountability, and transparency

    Ustun, B., Spangher, A., Liu, Y.: Actionable recourse in linear classification. In: Proceedings of the conference on fairness, accountability, and transparency. pp. 10–19 (2019) Pareto Optimal Algorithmic Recourse in Multi-cost Function XV

  31. [39]

    In: Proceedings of the 2020 conference on fairness, accountability, and transparency

    Venkatasubramanian, S., Alfano, M.: The philosophical basis of algorithmic re- course. In: Proceedings of the 2020 conference on fairness, accountability, and transparency. pp. 284–293 (2020)

  32. [40]

    arXiv preprint arXiv:2010.10596 (2020)

    Verma, S., Boonsanong, V., Hoang, M., Hines, K.E., Dickerson, J.P., Shah, C.: Counterfactual explanations and algorithmic recourses for machine learning: A review. arXiv preprint arXiv:2010.10596 (2020)

  33. [41]

    Neurocomputing429, 215– 244 (2021)

    Wang, M., Deng, W.: Deep face recognition: A survey. Neurocomputing429, 215– 244 (2021)

  34. [42]

    Complex & Intelligent Systems7, 439–457 (2021)

    Zhang, Q., Lu, J., Jin, Y.: Artificial intelligence in recommender systems. Complex & Intelligent Systems7, 439–457 (2021)

  35. [43]

    IEEE transactions on neural networks and learning systems30(11), 3212– 3232 (2019)

    Zhao, Z.Q., Zheng, P., Xu, S.t., Wu, X.: Object detection with deep learning: A review. IEEE transactions on neural networks and learning systems30(11), 3212– 3232 (2019)

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.