acceptodds
Under review as a conference paper at ICLR 2027

A Robustified Greedy Algorithm for Online Transportation with Improved Competitive Guarantees

Abstract

We study the online transportation problem, in which requests arriving sequentially in a metric space must be irrevocably assigned to capacitated facilities. Beyond classical logistics applications, this problem models resource-allocation tasks arising in machine learning, including online facility assignments, recommender systems, and mixture-of-experts routing. We introduce Robustified Greedy (RG), a deterministic generalization of the Robust Matching algorithm that achieves a competitive ratio of , improving upon the state-of-the-art bounds of (Arndt et al., SOSA 2026) and (Harada and Itoh, ICALP 2025). RG also retains the metric-sensitive guarantee established for Robust Matching (RM) (Nayyar and Raghvendra, FOCS 2017), achieving a competitive ratio of in -dimensional Euclidean spaces for fixed . No comparable metric-sensitive guarantee is known for the transportation algorithms of Arndt et al. or Harada and Itoh. Beyond these competitive guarantees, RG provides a simple explanation for its decisions. It favors the natural nearest-neighbor assignment and, for suitable parameters, departs from this choice only when it identifies a reassignment that reduces the cost of its maintained auxiliary matching, thereby correcting accumulated assignment costs. We also prove that nearest-neighbor assignments account for a guaranteed fraction of RG's total cost, approaching one-half for appropriate parameters, even under adversarial arrivals. Experiments on real-world datasets corroborate the theory: RG achieves lower cost-to-OPT ratios than the competing algorithms while retaining a substantial nearest-neighbor component in its cost.

Primary area:
optimization
Keywords:
online algorithms, online transportation problem, greedy algorithms, online facility assignment

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

This paper studies the online metric transportation problem. There are kk server locations and nn servers distributed among these locations. There are nn requests arriving sequentially, each needed to be matched to an available server immediately and irrevocably. Then cost of a match is measured by the distance. The goal is to design algorithms whose competitive ratio only depends on kk.

The authors propose the Robustified Greedy (RG) algorithm, which generalizes the Robust Matching (RM) algorithm of (Raghvendra, 2016). They show that RG achieves a competitive ratio at most 6.6604k−2.896.6604k - 2.89, improving the previously best-known competitive ratio 8k−78k - 7. Moreover, they show that RG retains the metric-sensitive property of RM so that its competitive ratio adapts to the dimension of the Euclidean space. Finally, they lower bound the total fraction of online cost of RG from greedy assginments (i.e., assigning the arriving request to the nearest-available server) by 1/2−(α+1)/2(t−α)1/2 - (\alpha + 1) / 2(t - \alpha), where tt and α\alpha are tunable parameters of the algorithm. Experiments showcase that RG exhibits excellent empirical performance.

Critical strengths

This paper considers a well-motivated and widely-studied problem. The writing is generally clear, and I appreciate authors' effort on presenting the results. The proofs are correct to the extent that I've checked. The proposed algorithm generalizes an existing algorithm in a natural (albeit less intuitive) way, improving the best competitive-ratio upper bound for this problem. To the best of my knowledge, the guarantee on the fraction of greedy assignments is new to the literature.

Critical weaknesses

The motivating question about greedy assignments is, in my opinion, not very appealing. It's unclear to me why we should caring about the greedy-assignment cost of the algorithm given that it's not related to the competitive ratio. In addition, the non-greedy assignments also affect the server configurations and hence the greedy assignments, making the greedy-assignment cost an algorithm-specific benchmark.

On the technical side, the primal-dual analysis of RG seems to substantially resemble the analysis of RM presented in (Raghvendra, 2016). Although I haven't checked the metric-sensitivity proof, it wouldn't be surprising to me that it follows the same proof template as the prior work. Overall, I don't see much new technical insight coming from the proofs.

There are other minor weakness, including the missing Related Work section and a clear exposition of proof intuition for those who never read (Raghvendra, 2016).

Questions for authors
  • What is the intuition behind reducing the penalty term for direct augmenting paths? Why would it lead to improved competitive ratio?

  • What are the main technical innovations compared to the analysis for RM in previous work?

  • What is the competitive ratio of RG for online metric matching?

3 comments

We sincerely thank the reviewer for their feedback. We clarify several aspects of our contributions that may not have been sufficiently clear and will revise the paper accordingly.


Reviewer concern. It's unclear to me why we should caring about the greedy-assignment cost of the algorithm given that it's not related to the competitive ratio.

Response. Greedy assignments are central to achieving RG's improved competitive ratio of 6.6604k−2.896.6604k-2.89. We explain this connection through our technical contributions below.



Reviewer concern. On the technical side, the primal-dual analysis of RG seems to substantially resemble the analysis of RM presented in (Raghvendra, 2016). [...] Overall, I don't see much new technical insight coming from the proofs.

Response. RG builds on RM's primal-dual framework. However, our work introduces three new algorithmic and analytical tools:

  • Saturation-epoch analysis. We partition the execution into epochs defined by server saturation and develop a new charging argument within each epoch. This technique establishes an O(k)O(k) competitive ratio for both RM and RG, independent of total server capacity. In particular, it yields an 11.66k11.66k bound for RM, improving upon the previously known O(klog⁡2n)O(k\log^2 n) guarantee.
  • Discounting direct augmenting paths. RG introduces a parameter α<t\alpha < t that discounts direct augmenting paths relative to non-direct paths. Whenever RG selects a direct path, it necessarily assigns the request to its nearest available server.
  • Persistence of greedy assignments. The epoch analysis alone yields the 11.66k11.66k bound. The improvement to 6.6604k−2.896.6604k-2.89 directly exploits RG's nearest-neighbor assignments: these assignments persist in the auxiliary matching until the corresponding server becomes saturated and the epoch ends (Lemma 10). This property is crucial to obtaining a tighter charging argument and the stronger competitive guarantee.


Reviewer concern. The motivating question about greedy assignments is, in my opinion, not very appealing.

Response. Greedy assignments have independent practical significance beyond their role in our competitive analysis. We discuss this motivation in our common response.



Reviewer question. What is the intuition behind reducing the penalty term for direct augmenting paths? Why would it lead to improved competitive ratio?

Response. Choosing α<t\alpha<t discounts direct edges from the arriving request to free servers relative to non-direct augmenting paths. Among all direct paths, the edge to the nearest free server has the lowest net-cost. Thus, whenever RG selects a direct path, it necessarily chooses the nearest available server. This greedy assignment enters both the online and auxiliary matchings and persists in the auxiliary matching until the corresponding server becomes saturated (Lemma 10). Exploiting this persistence enables a tighter analysis, improving the competitive-ratio bound from 11.66k11.66k to 6.6604k−2.896.6604k-2.89.



Reviewer question. What is the competitive ratio of RG for online metric matching?

Response. For the specific setting of online metric matching, a sharper analysis yields a sharper competitive ratio of 1+2(n−1)max⁡1 + 2(n-1)\max{$1,\frac{t}{\alpha+1}$} which matches the tight 2n−12n-1 for α≥t−1\alpha \geq t-1.

AuthorsOpenReview ↗

response request

Thank you again for your review. With the discussion closing tonight, we’d appreciate a brief acknowledgment of our rebuttal and any feedback on whether it addresses your concerns.

Dear authors,

I really appreciate your efforts on addressing my concerns. I'm still not convinced by the value of measuring the contribution of Greedy assignments since the value of Greedy mainly comes from its easy implementation and transparency, yet the proposed algorithm still admits rather complicated form. On the technical side, improving the leading constant seems quite marginal to me. Overall, I'd like to maintain my assessment and score.

Reviewer_sMd1OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes Robustified Greedy (RG), a deterministic algorithm for online transportation that extends Robust Matching by favoring nearest-available-server assignments. Using an analysis based on server saturation epochs, it improves the competitive upper bound from 8k−78k-7 to approximately 6.6604k−2.896.6604k-2.89, where kk is the number of server locations. It also establishes geometry-sensitive guarantees and a lower bound on the cost fraction contributed by greedy assignments. Experiments on NYC Taxi and MovieLens evaluate assignment cost, runtime, and nearest-neighbor behavior.

Critical strengths

The paper advances the best-known competitive upper bound for deterministic online transportation in general metric spaces, improving it from 8k−78k-7 to approximately 6.6604k−2.896.6604k-2.89, independently of the total capacity.

Critical weaknesses
  1. The paper improves a constant within the existing O(k)O(k) deterministic guarantee, while randomized algorithms already achieve O(log⁡2k)O(\log^2 k) expected competitiveness against an oblivious adversary. The importance of this incremental result depends on a stronger motivation for determinism or broadly applicable technical insights, neither of which is sufficiently established.

  2. The empirical advantage is not convincing. RG largely matches Greedy on natural arrivals, offers only marginal gains on MovieLens, and is outperformed by RM on the NYC directional benchmark.

Questions for authors

See Critical Weaknesses. Consider modifying the experimental figures to better illustrate the advantages of your algorithm.

2 comments

We thank the reviewer for their feedback. We address two central concerns.


Reviewer concern. Randomized algorithms already achieve O(log⁡2k)O(\log^2 k) expected competitiveness against an oblivious adversary.

Response. The cited randomized algorithms have two important limitations: (1) their guarantees do not extend to adaptive adversaries, whose practical relevance is well documented, and (2) their strong theoretical guarantees do not necessarily translate into good empirical performance, even on benign inputs.

(1) Why are adaptive adversaries relevant in practice? The cited randomized guarantees hold against oblivious adversaries, not adaptive adversaries that observe previous decisions before choosing their next move. Such adaptive behavior has been documented in real-world matching markets, including strategic manipulation of ride-hailing platforms (Tripathy et al., Decision Sciences 2023; Cameron, Adm. Sci. Q. 2024). Adaptive adversaries are also extensively studied in online learning (Arora et al., ICML 2012; Cesa-Bianchi et al., NeurIPS 2013). Thus, robustness against adaptive adversaries is a practically motivated objective, not merely a preference for determinism.

(2) Why are existing randomized algorithms insufficient in practice? Tong et al. (PVLDB 2016) demonstrated Greedy's substantial empirical advantage over HST-based algorithms across more than 50,000 experiments. In Table 4 (Tong et al., PVLDB 2016), HST-Reassignment incurs 120–151% higher matching cost than Greedy, even on normal and uniform distributions. We suspect that the poor performance of HST-based approaches reflects errors introduced at three stages: metric embedding, nearest-neighbor mapping, and online matching on the tree.



Reviewer concern. The empirical advantage is not convincing. RG largely matches Greedy on natural arrivals, offers only marginal gains on MovieLens, and is outperformed by RM on the NYC directional benchmark.

Response. These comparisons overlook RG's distinct objectives. Our goal is not to outperform Greedy on benign inputs, but to preserve its strong empirical performance while providing adversarial robustness. Similarly, our goal is not to outperform RM on empirical cost, but to favor more explainable nearest-neighbor assignments. Section 5 evaluates both objectives:

  • RG versus Greedy (robustness). On natural NYC arrivals (Figure 1(a)), RG stays within 0.5% of Greedy's cost, preserving its strong empirical performance. On directional arrivals (Figures 1(b)–(c)), RG improves upon Greedy by up to 15.02%, with gains increasing with the number of server locations. Please see Table 1 in the common response section for a summary of the result and Figure 2 in Section 5 of the modified paper for more details.

  • RG versus RM (explainability). RG makes 97.37% nearest-neighbor assignments versus RM's 93.44% on NYC, and 63.76% versus 47.40% on MovieLens (Section 5). We have also added the results for the newly conducted adversarial test in Section 5 and provide a brief summary in Table 2 of the common response section. Thus, RG preserves more of Greedy's explainable decisions while achieving a stronger theoretical competitive guarantee than RM. This is precisely the trade-off our experiments are designed to demonstrate.

AuthorsOpenReview ↗

response request

Thank you again for your review. With the discussion closing tonight, we’d appreciate a brief acknowledgment of our rebuttal and any feedback on whether it addresses your concerns.

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

The paper studies online metric transportation: given a metric space (X,d)(X, d), with kk special locations SS (called servers). Each server s∈Ss \in S has some positive integer capacity bsb_s with total capacity nn for given integer nn; that is, ∑s∈Sbs=n\sum_{s \in S} b_s = n. Exactly nn requests arrive online one-by-one, each at a point xi∈X,i∈[n]x_i \in X, i \in [n] in the metric space. Each request must be assigned irrevocably to one of the k:=∣S∣k := |S| servers when it arrives. The objective is to minimize the total assignment distance, and the competitive ratio of an algorithm is defined as the worst-case ratio of its total assignment distance to the best possible total assignment distance in hindsight, where the worst-case is over all possible instances. This generalizes the online metric matching problem.

The paper introduces Robust Greedy (RG), a two-parameter generalization of Robust Matching (RM) that discounts direct augmenting paths, thereby favoring assignments to the nearest available server.

The main contribution is a competitive-ratio bound of 6.6604k−2.896.6604k - 2.89, where kk is the number of server locations, improving the leading constant in the previous 8k−78k - 7 bound. The analysis also establishes an O(k)O(k) guarantee for RM and extends its geometry-sensitive guarantees to RG. In a separate parameter regime, the paper proves that nearest-neighbor assignments contribute a guaranteed fraction of the online cost.

Experiments on NYC taxi and MovieLens data compare RG against Greedy, RM, and the algorithm of Arndt et al. However, results are quite mixed here. While their algorithm (RG) outperforms Greedy in one setting, it does not outperform RM.

Critical strengths

The main strength of the paper is the algorithm and its analysis. The algorithm is clean: it extends Robust Matching by introducing a parameter that discounts direct augmenting paths, thereby favoring nearest-neighbor assignments. The analysis is nontrivial but remains relatively concise and transparent. In particular, partitioning the execution into server-saturation epochs and exploiting the persistence of direct assignments in the auxiliary matching provide useful structural insights.

The analysis improves the leading coefficient in the state-of-the-art competitive guarantee from 88 to 6.66046.6604, yielding a bound of 6.6604k−2.896.6604 k - 2.89, where kk is the number of server locations. The guarantee is independent of the total capacity of the servers. Moreover, the framework establishes an O(k)O(k) competitive ratio for the original Robust Matching algorithm, resolving a question identified in prior work.

The analysis looks correct to me.

Critical weaknesses

The experiments do not clearly establish an advantage over Robust Matching (RM). On NYC directional arrivals, RG is claimed to improve total cost over Greedy by at most 15.0%, whereas RM achieves a 24.3% improvement. Although RG makes more nearest-neighbor assignments, the paper does not establish the practical value of this property independently of total assignment cost, which is the stated objective. Consequently, the empirical case for preferring RG to RM remains unclear.

The improvements over Greedy are limited and depend on the experimental setting. RG essentially matches Greedy on NYC timestamp arrivals and on MovieLens with t=6t = 6, while the reported MovieLens gains at t=1.12t = 1.12 are only 0.41%-0.51%. In Figure 1(b), the highlighted 15.02% reduction concerns final cumulative cost after all 10,000 requests have consumed the available capacity; the advantage appears to emerge primarily near the end. This is a valid comparison under the paper's model, but it leaves unclear whether comparable benefits occur at lower capacity utilization, or if they can be sustained with more requests. Further, the paper provides limited. Further, the paper provides limited justification for the specific experimental parameter choices.

Some typos/suggestions:

  • Please change the claim of 6.66k - 0.89 in the abstract; use either the actual number (around 6.6604) or a number such as 6.67 or 20/3 that is at least as large as the actual number.
  • The x-axis of each plot in Figure 1 use "k" to shorter the numbers. This is slightly confusing at first glance since k also denotes the number of servers. Instead the x-axis label could be "number of requests (in multiple of 1000)".
Questions for authors
  1. Do the experiments use the primal-dual implementation in Section 3? If not, could that make it faster?
  2. Do the experiments sweep over all possible values, choosing the best ones in hindsight? The rationale for the specific values should be made explicit, even if the choices were arbitrary.
  3. Could you add error bars on the plots?
  4. Do you have a sense of whether your RG algorithm (with optimal parameters or otherwise) has a lower bound?
1 comment

We thank the reviewer for recognizing our theoretical contributions and for a positive evaluation of our paper. We address the critical weaknesses first.


The experiments do not clearly establish an advantage over Robust Matching (RM)[...] Consequently, the empirical case for preferring RG to RM remains unclear.

The improvements over Greedy are limited and depend on the experimental setting. RG essentially matches Greedy on NYC timestamp arrivals and on MovieLens with t=6t=6, while the reported MovieLens gains at t=1.12t=1.12 are only 0.41%-0.51%.

Response. As stated in the common response, Greedy is a gold standard for online matching, both for its strong empirical performance (Tong et al., PVLDB 2016) and for the simplicity and explainability of its decisions (Balkanski et al., EC 2023; Yang and Yu, 2026). However, it is vulnerable to adaptive adversarial inputs. The goal of our experiments is not to establish that RG outperforms Greedy on benign inputs, but to show that it preserves Greedy's strong empirical performance while providing adversarial robustness. Similarly, our goal is not to outperform RM on empirical cost, but to favor more explainable nearest-neighbor assignments. We have reorganized Section 5 to emphasize the following objectives.

  • RG versus Greedy (robustness). On natural NYC arrivals (Figure 1(a)), RG stays within 0.5% of Greedy's cost, preserving its strong empirical performance. On directional arrivals (Figures 1(b)–(c)), RG improves upon Greedy by up to 15.02%, with gains increasing with the number of server locations. We have added a stronger experiment using adversarially adaptive request sequences across Greedy, RG and RM. We observe that Greedy can be as much as 17x higher in terms of competitive ratio compared to RG in this case. Please see Table 1 in the common response section for a summary of the result.
  • RG versus RM (explainability). RG makes 97.37% nearest-neighbor assignments versus RM's 93.44% on NYC, and 63.76% versus 47.40% on MovieLens (Section 5). We have also added the results for the newly conducted adversarial test in Section 5 and provide a brief summary in Table 2 of the common response section. Across the tested settings, RG remains comparable to RM in assignment cost or observed competitive ratio while assigning a larger fraction of requests to the nearest available server, making more of its decisions explainable by a simple local rule. This is precisely the trade-off our experiments are designed to demonstrate.


In Figure 1(b), the highlighted 15.02% reduction concerns final cumulative cost after all 10,000 requests have consumed the available capacity; the advantage appears to emerge primarily near the end [...] or if they can be sustained with more requests.

Response. It is indeed the case that the maximum benefits of RG over Greedy can be seen towards the end of the sequence when server locations are nearly saturated and this is an expected behavior in natural settings. However, our newly conducted adversarial experiments showcase substantial gain of RG over Greedy even at the earlier part of the sequence (Figure 2 of updated paper).



Further, the paper provides limited justification for the specific experimental parameter choices.

Response. For our experiments, the parameter choices are mostly arbitrary. We chose α∈1,1.5,2\alpha\in{1,1.5,2}. For these values of α\alpha, t=6t=6 is a natural choice, as it is the smallest integer satisfying the condition t>5t>5 required by Lemma 6. We use this value for NYC-Taxi and Synthetic, including the newly added adaptive-adversary experiment. For MovieLens, however, the most visually distinguishable results were obtained empirically at t=1.12t=1.12 and α∈1.02,1.05,1.08,1.10\alpha\in{1.02,1.05,1.08,1.10}.



Some typos/suggestions:

Response. Thanks for these suggestions. We have updated the paper with both suggestions.



Additional Questions

  1. Do the experiments use the primal-dual implementation in Section 3? If not, could that make it faster?

Response. Yes, we have used a highly optimized version (to the best of our efforts) of the primal-dual implementation. In particular, we had achieved significant speedup using the additional properties like augmenting paths can consist of at most 2k2k hops.

2. Do the experiments sweep over all possible values, choosing the best ones in hindsight? The rationale for the specific values should be made explicit, even if the choices were arbitrary.

Response. Addressed above.

3. Could you add error bars on the plots?

Response. Thanks for the suggestion. We have added error bars in the updated version of the paper.

4. Do you have a sense of whether your RG algorithm (with optimal parameters or otherwise) has a lower bound?

Response. Despite a significant amount of effort, we did not find any lower bound for either RG or RM that is stronger than the standard 2k-1 lower bound for all deterministic algorithms.

Comments on the paper

1 comment

We thank the reviewers for their feedback. We address two central points: (1) the theoretical and practical benefits of preserving greedy assignments, and (2) the significance of our experimental results.

(1) Why robustify Greedy? Greedy assigns each request to its nearest available server, making its decisions easy to explain (Balkanski et al., EC 2023). It is also an empirical gold standard, outperforming online algorithms such as Permutation and HST-based methods in nearly all of the 50,000+ experiments of Tong et al. (PVLDB 2016). Greedy is also deployed in practice, e.g., GrabShare (Widdows et al., ICITE 2017). However, Greedy has poor worst-case guarantees: its competitive ratio under adversarial arrivals can be 2n−12^n-1, and Charikar et al. (2026, STOC 2027 submission) recently established an Ω(n/log⁡n)\Omega(n/\sqrt{\log n}) lower bound even under random arrival order. Prior work seeks assumptions under which Greedy is competitive (Balkanski et al., EC 2023; Yang and Yu, 2026). We instead ask (Introduction, lines 99–103): Can we favor greedy assignments while guaranteeing strong worst-case performance?

  • Explainable choices. RG maintains an auxiliary matching to guide its decisions, and each nearest-neighbor assignment enters both the online and auxiliary matchings. RG deviates from Greedy only when an augmenting path can significantly reduce the cost of the auxiliary matching (Lemma 5). Thus, even RG's non-greedy choices are explainable: adversarial inputs can make greedy assignments accumulate high costs, which the auxiliary matching reflects, and RG makes non-greedy choices precisely to mitigate these effects. Moreover, a significant portion of RG's online cost comes from nearest-neighbor assignments (Lemma 6).
  • Improved robustness. For the online transportation problem, RG achieves a competitive ratio of 6.6604k−2.896.6604k-2.89, improving on our bound of 11.66k11.66k for RM. This improvement directly exploits the structure of the auxiliary matching: nearest-neighbor assignments persist in this matching until a server becomes fully saturated and the corresponding epoch ends (Lemma 10). This persistence enables a tighter analysis and a stronger competitive guarantee.

(2) Experimental Results Our experimental goal is therefore not to outperform Greedy on natural data, but to highlight RG's robustness to adversarial inputs. Likewise, our goal is not to outperform RM, but to show that RG makes more nearest-neighbor assignments and thus more explainable decisions.

  • RG versus Greedy — robustness. RG lowers cost by up to 15.02% on the submitted NYC directional-arrival test. We have added a stronger experiment using adversarially adaptive request sequences on Greedy, RG and RM, where the competitive ratio of RG is as much as 17x lower than that of Greedy (Figure 2 in Section 5 of the updated paper). Table 1 summarizes this result.

Table 1

Algorithmα\alphaCost/OPT
Greedy–6,837.07
RM–437.28
RG1391.89
RG1.5383.02
RG2400.83
  • RG versus RM — explainability. We measure nearest-neighbor (NN) decisions alongside total cost. RG makes 97.37% versus RM's 93.44% on NYC, and 63.76% versus 47.40% on MovieLens, under the parameter settings reported in the paper. Table 2 shows the same for the new adversarial test (also in Section 5 of the updated paper). RG makes significantly more NN decisions than RM while being comparable in competitive ratio (Table 1).

Table 2

Algorithmα\alphaNN decisionsNN cost share
Greedy–––
RM–67.95%35.00%
RG191.17%60.14%
RG1.590.60%59.19%
RG289.73%57.43%

References

[1] Raman Arora, Ofer Dekel, and Ambuj Tewari. Online Bandit Learning against an Adaptive Adversary: From Regret to Policy Regret. In Proceedings of the 29th International Conference on Machine Learning (ICML), pp. 1503–1510, 2012.

[2] Lindsey D. Cameron. The Making of the “Good Bad” Job: How Algorithmic Management Manufactures Consent Through Constant and Confined Choices. Administrative Science Quarterly, 69(2):458–514, 2024.

[3] Nicolò Cesa-Bianchi, Ofer Dekel, and Ohad Shamir. Online Learning with Switching Costs and Other Adaptive Adversaries. In Advances in Neural Information Processing Systems 26 (NIPS), pp. 1160–1168, 2013.

[4] Moses Charikar, Alireza Haqi, Jabari Hastings, Yichen Huang, and Prasanna Ramakrishnan. Sublinear Metric Distortion for Matching. arXiv preprint arXiv:2611.05777, 2026.

[5] Manish Tripathy, Jiaru Bai, and H. Sebastian Heese. Driver Collusion in Ride-Hailing Platforms. Decision Sciences, 54(4):434–446, 2023.

[6] Dominic Widdows, Jacob Lucas, Muchen Tang, and Weilun Wu. GrabShare: The Construction of a Realtime Ridesharing Service. In 2017 2nd IEEE International Conference on Intelligent Transportation Engineering (ICITE), pp. 138–143, 2017.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.