Submodular Optimization & Approximation Algorithms
Robustness and Generalization of Graph Submodular Optimization to Weighted, Directed, and Attributed Networks
Generated automatically from the limitations stated in 8 papers (TheWebConference, KDD, SIGIR), listed under Evidence. It is not a paper, and it does not come from papers submitted to CSPaper.
The problem
Current submodular optimization algorithms on graphs are almost exclusively formulated and evaluated on static, unweighted, and simple undirected graphs. Practitioners working with networks containing edge capacities, asymmetric directed relations, or node attributes have no evidence regarding whether these algorithms retain their approximation quality or degrade catastrophically. Without systematic testing across these graph variants, it remains unknown which algorithms generalize gracefully and which require fundamentally new theoretical formulations.
Why it matters
Reveals the exact failure modes and performance boundaries of existing submodular algorithms when applied beyond simple unweighted graphs. Provides researchers with an empirical baseline that indicates where novel theoretical mechanisms are strictly necessary versus where existing methods suffice.
Ways to approach it
Prior-work checks are free with an account. Results someone already ran are shown to everyone.
- 1
Empirical Robustness Benchmark: Collect open-source implementations of prominent submodular graph optimization algorithms and benchmark them across synthetic and real-world graphs augmented with edge weights, directionality, and node attributes, measuring empirical approximation ratios and runtime degradation relative to unweighted baselines.
- 2
Reduction and Adaptation Evaluation: Implement standard algorithmic adaptations (such as weighted marginal gain heuristics and bidirectional transformations) and measure their objective value retention against exact integer programming baselines on non-simple graph topologies.
Have a different approach?
Describe how you would tackle this problem and we'll look for papers that already do it. Free; your text stays private.
Why it might fail
If simple weight-scaling or symmetrization reductions readily preserve approximation guarantees across all evaluated problem classes, the robustness study will find little performance variation to analyze. Conversely, if each individual algorithm's failure mode is entirely idiosyncratic to its specific submodular objective, the findings will not generalize across the literature.
Evidence
Each paper's own statement of the limitation, verbatim.
- Scalable and Provable Biclique-Preserving Clustering: The Power of Counting-based ApproachesTheWebConference 2026
Designed strictly for static, unweighted bipartite graphs and does not handle attributed or dynamic networks.
- Network Dismantling via Reverse Dismantling: Static and Dynamic AlgorithmsTheWebConference 2026
Restricted to unweighted and undirected graphs, lacking mechanisms for weighted node costs or edge capacities.
- Maximum Edge-based Quasi-Clique: Novel Iterative FrameworksTheWebConference 2026
Limited to unweighted, simple undirected graphs; extension to weighted or directed networks is unaddressed.
- Fast Algorithms for Group Markov Centrality OptimizationKDD 2026
Formulation and theoretical guarantees are restricted to unweighted undirected connected graphs.
- Cohesive Group Discovery in Interaction Graphs under Explicit Density ConstraintsSIGIR 2026
Restricted to unweighted and unattributed simple graphs, lacking support for edge weights or node features.
- Accelerated Coordinate Descent for Directed Densest Subgraph DiscoveryKDD 2026
Evaluated and formulated strictly for static, unweighted directed graphs.
- Highly-efficient minimization of network connectivity in large-scale graphsTheWebConference 2025
Evaluated only on undirected, unweighted graphs; the O(m) space for storing A still applies, and directed/weighted settings are not addressed
- Learning to Sample Hard Instances for Graph AlgorithmsACML 2019
Current formulation is restricted to undirected, unweighted, and simple graphs.
Nearest existing work
- Submodular Maximization under 𝑘 -System Constraints in Parallel: A Trifecta of Approximation, Adaptivity, and Query ComplexityKDD 2026
- Structured Robust Submodular Maximization: Offline and Online AlgorithmsAISTATS 2019
- Submodular Maximization beyond Non-negativity: Guarantees, Fast Algorithms, and ApplicationsICML 2019
- Robust Guarantees of Stochastic Greedy AlgorithmsICML 2017
- Filtered Search for Submodular Maximization with Controllable Approximation BoundsAISTATS 2015
- An Asymptotically Optimal Approximation Algorithm for Multiobjective Submodular Maximization at ScaleICML 2025
- Robustness in Multi-Objective Submodular Optimization: a Quantile ApproachICML 2022
- Instance Specific Approximations for Submodular MaximizationICML 2021
- Robust Submodular Maximization: A Non-Uniform Partitioning ApproachICML 2017
- Maximization of Monotone $k$-Submodular Functions with Bounded Curvature and Non-$k$-Submodular FunctionsACML 2021
- Maximization of Monotone $k$-Submodular Functions with Bounded Curvature and Non-$k$-Submodular FunctionsACML 2021
- Fast Adaptive Non-Monotone Submodular Maximization Subject to a Knapsack ConstraintNeurIPS 2020
- The Power of Randomization: Distributed Submodular Maximization on Massive DatasetsICML 2015
- Randomized Algorithms for Submodular Function Maximization with a $k$-System ConstraintICML 2021
- Submodular Cost Submodular Cover with an Approximate OracleICML 2019