Data-Driven Solution Portfolios

Andreas Maggiori; Marina Drygala; Miltiadis Stouras; Ola Svensson; Sergei Vassilvitskii; Silvio Lattanzi

REVIEW 2 major objections 3 minor 3 cited by

Data-Driven Solution Portfolios

T0 review · 2 major / 3 minor · reviewed 2026-08-12 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A polynomial-time algorithm constructs a portfolio of k matroid solutions whose best is within a constant factor of the optimum.

desk verdict Fresh problem, clever algorithm, but a real gap in Lemma 5.5 makes the main theorem unproven as written. read the letter →

arxiv 2412.00717 v1 pith:JHNHQ3FN submitted 2024-12-01 cs.DS

Marina Drygala , Silvio Lattanzi , Andreas Maggiori , Miltiadis Stouras , Ola Svensson , Sergei Vassilvitskii This is my paper · ORCID

classification cs.DS MSC 68W2505B3590C27

keywords portfoliooptimizationmatroidapproximationalgorithmcontentionresolutionschemestochasticdata-drivendesignPoissonbinomialsubmodularity

topics P versus NP

open problems P versus NP

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

This paper proposes a stochastic portfolio-optimization problem and claims an efficient solution for it. Given a matroid whose elements are independently active with known probabilities, the goal is to choose $k$ independent sets offline so that the expected size of the best one (the one with most active elements) is maximized. Because the solution family is usually exponential, the natural submodular greedy approach is infeasible. The paper's central claim is a polynomial-time algorithm returning a portfolio whose expected value is within a constant factor $\Theta(1)$ of the optimal portfolio, for every matroid. The significance is that a single, data-independent selection of $k$ solutions can hedge against uncertainty whenever the underlying feasibility constraint is a matroid.

What carries the argument

The central machinery is a column decomposition plus a feasible-sampling contention-resolution scheme (CRS). For a prefix of $d$ bases and next base $B_{d+1}$, basis-exchange bijections map each element of $B_{d+1}$ to $d$ elements in the prefix, producing columns; the column portfolio samples one element per column. The CRS used is designed for 'feasible sampling strategies' (Definition 5.1): marginal inclusion probabilities whose half-vector lies in the matroid polytope, plus a negative-dependence condition. Theorem 5.2 gives that every element is retained by the CRS with probability at least $1/8$, and the proof of the constant factor hinges on showing that even after conditioning on outlier events (the winning index $i^*$, the active columns $C^*$, or the sampled independent set $I^*$), the sampling remains feasible via deferred decisions. The remaining analysis uses stochastic dominance lemmas for Poisson-Binomial variables and a balls-and-bins bound on distinct sampled elements.

What would settle it

A direct numerical check would compute, on the smallest matroid where the column decomposition creates dependencies (for instance a graphic matroid on three elements with prefix length $d=2$), the conditional marginal vector and both sides of the negative-dependence inequality after conditioning on $i^*=i'$ and $C^*=C'$; if either condition of Definition 5.1 fails, Theorem 5.2's retention bound of $1/8$ cannot hold and the constant-factor claim collapses. Alternatively, simulate Algorithm 2 on random graphic matroids with adversarial probabilities and compare the empirical approximation ratio; a ratio tending to zero would disprove the theorem.

Watch

Extended reading notes

Core claim

The central claim is Theorem 3.1: under a product distribution (each element active independently with probability $p_e$), Algorithm 2 is a $\Theta(1)$-approximation for the $k$-portfolio solution problem over any matroid, with polynomial time complexity. The proof builds an ordering of disjoint highest-expectation bases, examines every prefix, and for each prefix constructs two portfolios — one by uniform sampling from the prefix, one by sampling one element from each 'column' of a basis-exchange decomposition — then trims each sampled set with a contention-resolution scheme. A key structural insight is that the optimal portfolio restricted to elements outside the chosen prefix already achieves $\Theta(1)$ of OPT, and these solutions can be stochastically dominated by independent copies of the next highest-expectation basis. The analysis then splits on whether high- or low-probability elements of that basis drive the optimum, showing the column portfolio handles the first case and the uniform portfolio the second. The paper also establishes that for general distributions no $(1-1/e+\epsilon)$-approximation is possible unless P=NP.

Load-bearing premise

The argument depends on the deferred-decisions claim that, after conditioning the sampling on events such as the winning solution $i^* = i'$ and the set of columns it hit, the conditional process still satisfies Definition 5.1 — that is, the scaled marginals stay in the matroid polytope and sampling an element does not raise its chance of being spanned by the other sampled elements.

Editorial extensions

If this is right

For uniform matroids, sorting by activation probability and trying prefixes gives a $\Theta(1)$-approximation; the paper's example shows a disjoint-only strategy can be exponentially worse, achieving $O(\log\log k)$ where mixing achieves $\Theta(\log k/\log\log k)$.
For arbitrary matroids, the algorithm runs in time polynomial in the ground set and portfolio size, so it applies to implicit solution families like spanning trees where listing all solutions is impossible.
For general (non-product) distributions, no polynomial-time algorithm beats $(1-1/e+\epsilon)$ unless P=NP, so the product-distribution assumption is essential.
Because the optimal value is submodular, the known greedy bound is $(1-1/e)$ but with running time polynomial in the number of solutions; the new result trades a constant-factor loss for polynomial dependence on the input description.

Reading between the lines

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

One testable extension is whether the same constant-factor guarantee survives positively correlated activation (for example, a common latent factor), since the proof's independence assumptions enter through the CRS feasibility and the Poisson-Binomial dominance lemmas.
The column-decomposition idea suggests a template for other implicitly defined solution sets: whenever basis-exchange bijections exist and a feasible sampling strategy can be defined, a similar prefix-and-portfolio algorithm may yield constant-factor guarantees.
The paper leaves open the tightest constant; reading off the proof constants suggests improvements are possible by sharpening the CRS retention probability or the concentration arguments, so the true approximability might be a smaller constant than the one exhibited.
An immediate practical heuristic suggested by the analysis is to test the empirical approximation on real traffic or sports-betting data, where the product assumption is only approximate; graceful degradation would indicate robustness.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Bottom line: this is a genuinely new algorithmic result with a clever algorithm, but the proof of the main theorem has a gap in Lemma 5.5 that needs to be fixed before the result is citable.

The paper formulates a natural problem—choosing k solutions to maximize the expected best under independent Bernoulli activations on a matroid—and shows the naive greedy/submodular view is not enough because the solution space is huge. The uniform-matroid warm-up (Section 4) is solid and clearly explained. The general algorithm, based on an ordering of disjoint bases, a column decomposition, and a contention resolution scheme, is interesting and the high-level strategy is sound. The paper also does the right thing in stating assumptions and proving they are w.l.o.g. There is no circularity or fitted constants.

The problem is in Lemma 5.5, Case 1 of the general analysis. The proof's step (5) assumes that, after conditioning on the identity of the winning set i*, the events e∈S_i* and e∈~A are independent. They are not. Conditioning on i* correlates the composition of the winning set with the activation pattern: a set is more likely to win if it sampled active elements. The stress-test's U_{2,4} example demonstrates this concretely: P(x∈V1,A_x=1|i*=1)=2/7, whereas the product of marginals is 1/4. So the factorization used to lower-bound the kept active elements is false. This is not a cosmetic issue: the 1/8 retention bound in Case 1 depends on it, and Case 1 is half of the main theorem. The proof needs an alternative argument—perhaps a positive-correlation inequality, since the bias seems to go in the right direction, or a different conditioning—to recover the bound. The minimal example does not suggest the theorem is false; the inequality may still hold. But as written, the main result is not fully proven for Case 1.

The rest of the analysis, including Lemma 5.7, has similar conditioning-on-outlier concerns, but I did not find a concrete counterexample there. The proof's deferred-decision arguments are plausible, just under-specified.

Who should read this: anyone working on data-driven algorithm design, portfolio optimization, or matroid contention resolution. The problem statement alone is worth knowing. But a serious referee should push for a corrected Lemma 5.5 before accepting. I would send it to review, and I hope the authors can fix the gap; if they do, the paper is a strong accept.

Referee Report

2 major / 3 minor

Summary. The paper introduces a stochastic portfolio-optimization problem over matroids: given a matroid and a product distribution over active elements, choose k independent sets so as to maximize the expected size of the best (largest) selected set under a random activation pattern. The main claim (Theorem 3.1) is a polynomial-time constant-factor approximation algorithm. The algorithm greedily orders disjoint maximum-expectation bases, considers all prefixes of this ordering, and for each prefix constructs two portfolios (uniform sampling and column-wise sampling), each followed by a contention resolution scheme. The analysis splits into a high-probability case, handled by the column portfolio, and a low-probability case, handled by the uniform portfolio. A simpler constant-factor algorithm for uniform matroids is given first as a warm-up.

Significance. The problem is well motivated and the paper contains a substantial algorithmic idea: constructing portfolios that exploit anti-concentration by mixing over columns, and using contention resolution schemes under conditioning on outlier events. The manuscript is self-contained, makes no use of fitted constants, and includes explicit reductions justifying the three simplifying assumptions, plus an NP-hardness reduction for general distributions. If the missing step in the matroid analysis is repaired, the paper would be a solid contribution to data-driven algorithm design and randomized combinatorial optimization. In its current form, however, one load-bearing step in the proof of the constant-factor guarantee is not justified.

major comments (2)

[5.4.1, Lemma 5.5, Eq. (5)] The proof of Lemma 5.5 passes from Eq. (4) to Eq. (5) by asserting that, conditioned on i*=i', the events e∈S_{i*} and e∈~A are independent, and therefore Pr[e∈S_{i*}∩~A | i*=i'] = Pr[e∈S_{i*}|i*=i'] · Pr[e∈~A|i*=i']. This is false: the conditioning event i*=argmax_i |V_i∩~A| correlates the sampled positions V_i with the activation pattern ~A, and this dependence propagates to S_{i*} because S_{i*} is obtained from V_{i*}. For a concrete instance, take the rank-2 uniform matroid U_{2,4}, prefix bases {x,z} and {y,w}, one column {x,y} with d=2, k=2, p_x=p_y=1/2, z,w inactive, and tie-breaking to the smaller index. Enumerating the sixteen equiprobable outcomes gives Pr[x∈V_1|i*=1]=7/13, Pr[A_x=1|i*=1]=6/13, but Pr[x∈V_1, A_x=1|i*=1]=4/13; the product is 42/169, not 4/13. The same failure persists with an idealized CRS that retains sampled elements independently with probability 1/8: conditional on i*=1, Pr[x∈S_1|i*=1]=(1/8)·(7/13) and Pr[A_x=1|i*=1]=6/13, but Pr[x∈S_1, A_x=1|i*=1]=(1/8)·(4/13), so the factorization used in line (5) is still invalid. Consequently, the proof of Lemma 5.5, and with it the Case 1 half of Theorem 3.1, is not established as written.
[5.4.1, Lemma 5.5, conditional feasibility] Lemma 5.5 also relies on the assertion that, after conditioning on i*=i' and C*=C', the sets V_i are still generated by a feasible sampling strategy in the sense of Definition 5.1. The proof gives a short argument for the marginal condition but only states that the negative-dependence condition 'can be proven by the principle of deferred decisions' without providing the actual argument. This is not a purely formal gap: conditioning on the argmax correlates the joint distribution of sampled elements with the activation pattern, so the inequality Pr[e∈span(V_i\{e}) | e∈V_i, i*, C*] ≤ Pr[e∈span(V_i\{e}) | i*, C*] needs a separate, non-obvious proof. As written, Theorem 5.2 cannot be invoked after this conditioning, so the 1/8 retention guarantee in Eq. (≀) is unsupported.

minor comments (3)

[Appendix C, Algorithm 5] The loop 'for i = 1, . . . , n do Order[n−i−1] = e*' appears to be off by one: for i=n it writes to index −1 and the last position Order[n−1] is never assigned. The intended assignment is probably Order[n−i] (or the loop should use zero-based indices consistently).
[Lemma 5.10] In the final display of the proof, the text says 'for (4) we used a union-bound', but the step replaces sum_{A'∈W} Pr[A=A'] by Pr[A∈W], which is an equality because the events are disjoint; the explanation should be corrected accordingly.
[Section 4, Algorithm 1] The pseudocode samples r elements with replacement and then removes duplicates, so the resulting sets may have size strictly less than r. The subsequent analysis treats S_i as a set of distinct elements; this is consistent, but it would help to state explicitly that the value of a solution counts distinct elements, and that feasibility under the uniform matroid is preserved.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the constant-factor approximation is derived from matroid exchange properties, elementary probability, and an independently proved contention resolution scheme, with no fitted input renamed as a prediction.

full rationale

The paper is a self-contained theoretical derivation. Algorithm 2 and its analysis compare the constructed portfolio directly to the value of the optimal portfolio OPT, which is defined in Section 2.1 as the maximizer of the expected best solution; no data is fitted and no target quantity is fed back into the algorithm. The load-bearing lemmas are proved in the paper or cited from independent external sources: the bijective exchange property is Corollary A.2 from Brualdi, the (1-1/e) rank lower bound is Corollary A.3 from Chekuri-Vondrak-Zenklusen, and the concentration bounds are standard. The contention resolution scheme used in Theorem 5.2 is constructed and proved in Appendix C rather than imported from the authors' prior work. Assumptions 1-3 are justified with explicit reductions (duplicate zero-probability elements, a direct 1/4 solution, and Lemma A.1), so they are not circular definitions. The paper's self-citations (LSV23, DVW24, MV22) appear only in the related-work survey and in statements that data-driven algorithm design is an active area; they do not carry the proof. The skeptical objection to Lemma 5.5, concerning the factorization Pr[e in S_i* and e in ~A | i* = i'] = Pr[e in S_i* | i* = i'] Pr[e in ~A | i* = i'], targets a probabilistic independence claim; if valid, it is a correctness gap in the proof, not a reduction of the theorem to its own assumptions. Hence no circular step can be quoted, and the score is 0.

Assumptions & free parameters 0 free parameters · 5 assumptions · 0 invented entities

No fitted parameters. All constants in the proof are absolute (e.g., 1/8, 1/3200) and arise from concentration arguments, not data fitting. The problem is defined with a product distribution and matroid constraints; the only modeling input is the independence assumption.

assumptions (5)

standard math Matroid basis exchange bijection (Corollary A.2, Brualdi 1969)
Used to construct the column decomposition in Section 5.3 and the transformations in Lemma 5.4.
standard math Chernoff bounds for sums of independent Bernoulli variables (Corollary A.4)
Used for the high-probability prefix event in Lemma 4.3 and column event probabilities in Lemma 5.6.
standard math Concentration and expectation of matroid rank under product distributions (Corollary A.3 and A.5)
Corollary A.3 is used in Lemma 5.8 to bound expected rank; Corollary A.5 is used to lower-bound the probability of the good event W in Section 5.4.2.
domain assumption Elements are active independently (product distribution D)
This is the stochastic model defining Matroid Portfolio Optimization (Section 2.2). The proofs of Lemmas 4.4 and 5.5 rely on independence of activations across elements.
domain assumption Access to sampling from D and known activation probabilities p_e
Algorithm 2 estimates portfolio values by sampling from D (Section 5.1); the ordering of bases uses p_e as weights.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Data-Driven Solution Portfolios." pith.science (2026). https://pith.science/paper/JHNHQ3FN

@misc{pith2026241200717,
  author       = {Pith},
  title        = {Pith review of: Data-Driven Solution Portfolios},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JHNHQ3FN}},
  note         = {Machine review of arXiv:2412.00717}
}

abstract

In this paper, we consider a new problem of portfolio optimization using stochastic information. In a setting where there is some uncertainty, we ask how to best select $k$ potential solutions, with the goal of optimizing the value of the best solution. More formally, given a combinatorial problem $\Pi$, a set of value functions $V$ over the solutions of $\Pi$, and a distribution $D$ over $V$, our goal is to select $k$ solutions of $\Pi$ that maximize or minimize the expected value of the {\em best} of those solutions. For a simple example, consider the classic knapsack problem: given a universe of elements each with unit weight and a positive value, the task is to select $r$ elements maximizing the total value. Now suppose that each element's weight comes from a (known) distribution. How should we select $k$ different solutions so that one of them is likely to yield a high value? In this work, we tackle this basic problem, and generalize it to the setting where the underlying set system forms a matroid. On the technical side, it is clear that the candidate solutions we select must be diverse and anti-correlated; however, it is not clear how to do so efficiently. Our main result is a polynomial-time algorithm that constructs a portfolio within a constant factor of the optimal.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

TCS-BENCH: Benchmarking State-of-the-Art Generative AI Theoretical Computer Science Research Ability
cs.CL 2026-08 reject novelty 7.0 of 10

The paper introduces TCS-Bench, a 300-task proof-generation benchmark from top TCS papers, and reports frontier LLM accuracies from 30% to 68% using an automated verifier.
Computing Diverse and Nice Triangulations
cs.CG 2025-06 reject novelty 7.0 of 10

A polynomial-time approximation framework for diverse near-optimal triangulations is presented, but it implicitly assumes the NP-hard optimum quality is known.
Navigating the Social Welfare Frontier: Portfolios for Multi-objective Reinforcement Learning
cs.LG 2025-02 conditional novelty 5.0 of 10

An α-approximate portfolio of RL policies can cover all p-mean social welfare objectives for p ≤ 1 with size O(log κ / log(1/α)), and the paper gives algorithms and experiments for this.

Reference graph

Works this paper leans on

23 extracted references · 19 canonical work pages · cited by 3 Pith papers

[1]

Data-driven algorithm design, 2020

Maria-Florina Balcan. Data-driven algorithm design, 2020

work page 2020
[2]

Learning sparse combinatorial representations via two-stage submodular maximization

Eric Balkanski, Baharan Mirzasoleiman, Andreas Krause, and Yaron Singer. Learning sparse combinatorial representations via two-stage submodular maximization. In International Conference on Machine Learning , pages 2207--2216. PMLR, 2016

work page 2016
[3]

Comments on bases in dependence structures

Richard A Brualdi. Comments on bases in dependence structures. Bulletin of the Australian Mathematical Society , 1(2):161--167, 1969

work page 1969
[4]

Faster fundamental graph algorithms via learned predictions

Justin Chen, Sandeep Silwal, Ali Vakilian, and Fred Zhang. Faster fundamental graph algorithms via learned predictions. In International Conference on Machine Learning , pages 3583--3602. PMLR, 2022

work page 2022
[5]

Submodular function maximization via the multilinear relaxation and contention resolution schemes

Chandra Chekuri, Jan Vondr\' a k, and Rico Zenklusen. Submodular function maximization via the multilinear relaxation and contention resolution schemes. In Proceedings of the Forty-Third Annual ACM Symposium on Theory of Computing , STOC '11, page 783–792, New York, NY, USA, 2011. Association for Computing Machinery

work page 2011
[6]

The madness of multiple entries in march madness, 2024

Jeff Decary, David Bergman, Carlos Cardonha, Jason Imbrogno, and Andrea Lodi. The madness of multiple entries in march madness, 2024

work page 2024
[7]

Faster matchings via learned duals

Michael Dinitz, Sungjin Im, Thomas Lavastida, Benjamin Moseley, and Sergei Vassilvitskii. Faster matchings via learned duals. Advances in neural information processing systems , 34:10393--10406, 2021

work page 2021
[8]

Predictive flows for faster ford-fulkerson

Sami Davies, Benjamin Moseley, Sergei Vassilvitskii, and Yuyan Wang. Predictive flows for faster ford-fulkerson. In International Conference on Machine Learning , pages 7231--7248. PMLR, 2023

work page 2023

Show all 23 references

[9]

The outer limits of contention resolution on matroids and connections to the secretary problem

Shaddin Dughmi. The outer limits of contention resolution on matroids and connections to the secretary problem. In 47th International Colloquium on Automata, Languages, and Programming (ICALP 2020) . Schloss-Dagstuhl-Leibniz Zentrum f \"u r Informatik, 2020

2020
[10]

Matroid secretary is equivalent to contention resolution

Shaddin Dughmi. Matroid secretary is equivalent to contention resolution. Innovations in Theoretical Computer Science (ITCS) , 2022

2022
[11]

Warm-starting push-relabel

Sami Davies, Sergei Vassilvitskii, and Yuyan Wang. Warm-starting push-relabel. arXiv preprint arXiv:2405.18568 , 2024

2024 arXiv
[12]

A threshold of ln n for approximating set cover

Uriel Feige. A threshold of ln n for approximating set cover. Journal of the ACM (JACM) , 45(4):634--652, 1998

1998
[13]

Balancing notions of equity: Approximation algorithms for fair portfolio of solutions in combinatorial optimization

Swati Gupta, Jai Moondra, and Mohit Singh. Balancing notions of equity: Approximation algorithms for fair portfolio of solutions in combinatorial optimization. arXiv preprint arXiv:2311.03230 , 2023

2023 arXiv
[14]

Segmentation problems

Jon Kleinberg, Christos Papadimitriou, and Prabhakar Raghavan. Segmentation problems. Journal of the ACM (JACM) , 51(2):263--280, 2004

2004
[15]

Team performance with test scores

Jon Kleinberg and Maithra Raghu. Team performance with test scores. ACM Transactions on Economics and Computation (TEAC) , 6(3-4):1--26, 2018

2018
[16]

Speeding up bellman ford via minimum violation permutations

Silvio Lattanzi, Ola Svensson, and Sergei Vassilvitskii. Speeding up bellman ford via minimum violation permutations. In International Conference on Machine Learning , pages 18584--18598. PMLR, 2023

2023
[17]

Algorithms with predictions

Michael Mitzenmacher and Sergei Vassilvitskii. Algorithms with predictions. Communications of the ACM , 65(7):33--35, 2022

2022
[18]

An analysis of approximations for maximizing submodular set functions—i

George L Nemhauser, Laurence A Wolsey, and Marshall L Fisher. An analysis of approximations for maximizing submodular set functions—i. Mathematical programming , 14:265--294, 1978

1978
[19]

Submodular dominance and applications

Frederick Qiu and Sahil Singla. Submodular dominance and applications. In Approximation, Randomization, and Combinatorial Optimization. Algorithms and Techniques (APPROX/RANDOM 2022) . Schloss Dagstuhl-Leibniz-Zentrum f \"u r Informatik, 2022

2022
[20]

Combinatorial optimization under uncertainty: Probing and stopping-time algorithms

Sahil Singla. Combinatorial optimization under uncertainty: Probing and stopping-time algorithms. Unpublished doctoral dissertation, Carnegie Mellon University , 2018

2018
[21]

Probabilistic submodular maximization in sub-linear time

Serban Stan, Morteza Zadimoghaddam, Andreas Krause, and Amin Karbasi. Probabilistic submodular maximization in sub-linear time. In International Conference on Machine Learning , pages 3241--3250. PMLR, 2017

2017
[22]

Submodular function maximization via the multilinear relaxation and contention resolution schemes

Jan Vondr \'a k, Chandra Chekuri, and Rico Zenklusen. Submodular function maximization via the multilinear relaxation and contention resolution schemes. In Proceedings of the forty-third annual ACM symposium on Theory of computing , pages 783--792, 2011

2011
[23]

A note on concentration of submodular functions, 2010

Jan Vondrak. A note on concentration of submodular functions, 2010

2010

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.

[1] [1]

Data-driven algorithm design, 2020

Maria-Florina Balcan. Data-driven algorithm design, 2020

work page 2020

[2] [2]

Learning sparse combinatorial representations via two-stage submodular maximization

Eric Balkanski, Baharan Mirzasoleiman, Andreas Krause, and Yaron Singer. Learning sparse combinatorial representations via two-stage submodular maximization. In International Conference on Machine Learning , pages 2207--2216. PMLR, 2016

work page 2016

[3] [3]

Comments on bases in dependence structures

Richard A Brualdi. Comments on bases in dependence structures. Bulletin of the Australian Mathematical Society , 1(2):161--167, 1969

work page 1969

[4] [4]

Faster fundamental graph algorithms via learned predictions

Justin Chen, Sandeep Silwal, Ali Vakilian, and Fred Zhang. Faster fundamental graph algorithms via learned predictions. In International Conference on Machine Learning , pages 3583--3602. PMLR, 2022

work page 2022

[5] [5]

Submodular function maximization via the multilinear relaxation and contention resolution schemes

Chandra Chekuri, Jan Vondr\' a k, and Rico Zenklusen. Submodular function maximization via the multilinear relaxation and contention resolution schemes. In Proceedings of the Forty-Third Annual ACM Symposium on Theory of Computing , STOC '11, page 783–792, New York, NY, USA, 2011. Association for Computing Machinery

work page 2011

[6] [6]

The madness of multiple entries in march madness, 2024

Jeff Decary, David Bergman, Carlos Cardonha, Jason Imbrogno, and Andrea Lodi. The madness of multiple entries in march madness, 2024

work page 2024

[7] [7]

Faster matchings via learned duals

Michael Dinitz, Sungjin Im, Thomas Lavastida, Benjamin Moseley, and Sergei Vassilvitskii. Faster matchings via learned duals. Advances in neural information processing systems , 34:10393--10406, 2021

work page 2021

[8] [8]

Predictive flows for faster ford-fulkerson

Sami Davies, Benjamin Moseley, Sergei Vassilvitskii, and Yuyan Wang. Predictive flows for faster ford-fulkerson. In International Conference on Machine Learning , pages 7231--7248. PMLR, 2023

work page 2023