Pruned Submodularity Graph-Based Optimization

• The paper introduces a preprocessing step using graphs to prune elements, reducing problem size while preserving near-greedy approximation guarantees.
• It leverages a weighted directed graph to measure pairwise dependence and redundancy, enabling randomized pruning that cuts computational cost.
• Domain-specific implementations in summarization, billboard selection, and reinforcement learning demonstrate its scalability and practical impact.

Searching arXiv for the primary PSG/submodular sparsification paper and related pruned-submodularity-graph applications. arXiv search query: "Scaling Submodular Maximization via Pruned Submodularity Graphs" Pruned submodularity graph-based approach denotes a family of scalable procedures for submodular optimization in which a weighted directed graph over ground elements is used to identify redundancy before the main optimization stage. The foundational formulation constructs a submodularity graph whose edge weights encode pairwise dependence under a submodular objective, repeatedly prunes elements with small divergence from a sampled representative set, and then runs greedy on the reduced ground set; the resulting approximation retains the standard greedy form up to an additive loss (Zhou et al., 2016). Later work adopts the same pattern explicitly as a pruned submodularity graph for domain-specific problems such as influential billboard slot selection and reinforcement learning with submodular rewards (Ali et al., 2022).

1. Foundational optimization setting

The original setting is cardinality-constrained submodular maximization on a finite ground set $V$ of size $n$:

$\max_{S\subseteq V,\ |S|\leq k} f(S).$

The function $f$ is assumed to be normalized, non-decreasing, and submodular. Submodularity is expressed through diminishing returns: for the marginal gain

$f(v|A)=f(A\cup\{v\})-f(A),$

the gain of adding $v$ is larger for smaller conditioning sets and smaller for larger sets. This is the structural reason greedy is applicable, and the same assumption reappears in later application papers, including billboard influence maximization and submodular reward RL (Zhou et al., 2016).

The classical greedy algorithm achieves the standard $1-1/e$ guarantee, but its cost becomes prohibitive at scale because each greedy step reevaluates marginal gains for essentially all remaining elements and is sequential and memory heavy. The pruned submodularity graph-based approach addresses this bottleneck by inserting a pruning layer before greedy or policy optimization. A plausible implication is that the method is best understood not as a replacement for submodular optimization, but as a preprocessing and reduction mechanism that preserves high-value structure while shrinking the candidate universe.

2. Submodularity graph formalism

The core object is a weighted directed graph $G(V,E,w)$ over the ground elements. For each ordered pair $(u,v)$, the edge weight is defined as

$w_{uv}=f(v|u)-f(u|V\backslash u).$

The first term, $n$0, measures how much $n$1 can still contribute when $n$2 is retained. The second term, $n$3, is a lower bound on how important $n$4 is globally, because submodularity implies $n$5 for any $n$6. Small $n$7 therefore indicates that $n$8 is relatively redundant given $n$9 (Zhou et al., 2016).

The graph is asymmetric. A conditional variant is also defined,

$\max_{S\subseteq V,\ |S|\leq k} f(S).$0

to describe pairwise relevance after conditioning on a current selected set $\max_{S\subseteq V,\ |S|\leq k} f(S).$1. The basic pruning argument relies on the inequality

$\max_{S\subseteq V,\ |S|\leq k} f(S).$2

which upper-bounds the marginal gain of $\max_{S\subseteq V,\ |S|\leq k} f(S).$3 by the marginal gain of another retained element $\max_{S\subseteq V,\ |S|\leq k} f(S).$4 plus the graph weight from $\max_{S\subseteq V,\ |S|\leq k} f(S).$5 to $\max_{S\subseteq V,\ |S|\leq k} f(S).$6. This turns graph proximity into a bound on optimization loss.

A second structural property is the directed triangle inequality

$\max_{S\subseteq V,\ |S|\leq k} f(S).$7

This is what allows sparsification without computing all $\max_{S\subseteq V,\ |S|\leq k} f(S).$8 edges exhaustively. In the formal sparsification problem, the divergence of a node $\max_{S\subseteq V,\ |S|\leq k} f(S).$9 from a retained subset $f$0 is

$f$1

and one seeks a subset $f$2 such that as many discarded elements as possible have small divergence: $f$3 The function $f$4 is non-monotone submodular, so solving this optimization exactly would itself be expensive (Zhou et al., 2016).

3. Randomized pruning via submodular sparsification

The practical algorithm introduced for the foundational problem is Submodular Sparsification (SS). It begins with the full ground set $f$5, maintains an output set $f$6, and iterates while $f$7. In each round, it samples $f$8 probe nodes uniformly at random to form $f$9, removes them from the current pool and adds them to $f(v|A)=f(A\cup\{v\})-f(A),$0, computes for every remaining $f(v|A)=f(A\cup\{v\})-f(A),$1

$f(v|A)=f(A\cup\{v\})-f(A),$2

and then removes the top $f(v|A)=f(A\cup\{v\})-f(A),$3 elements having the smallest $f(v|A)=f(A\cup\{v\})-f(A),$4. When the loop ends, the leftover nodes are added to $f(v|A)=f(A\cup\{v\})-f(A),$5 (Zhou et al., 2016).

The retained fraction per iteration is $f(v|A)=f(A\cup\{v\})-f(A),$6, so the shrinkage is aggressive. With the default choice $f(v|A)=f(A\cup\{v\})-f(A),$7, the removed fraction is

$f(v|A)=f(A\cup\{v\})-f(A),$8

Because the set shrinks by a factor $f(v|A)=f(A\cup\{v\})-f(A),$9 each round, the number of rounds is logarithmic, about $v$0. The probe set has size $v$1, so the method avoids exhaustive pairwise comparison while still exploiting graph structure. This suggests that the algorithm trades exact redundancy resolution for a randomized coverage of “important regions” of the graph.

Later papers preserve the same template. In influential billboard slot selection, the graph nodes are billboard slots, the edge weight is

$v$2

and pruning removes a $v$3 fraction of slots with the smallest divergence after sampling $v$4 probes (Ali et al., 2022). In reinforcement learning, SGPO (Submodularity Graph-based Policy Optimization) builds a graph over visited states, computes

$v$5

and prunes the states with the smallest divergence before training on the retained set $v$6 (Anand et al., 18 Jul 2025).

4. Approximation guarantees and computational profile

The theoretical guarantee for ideal sparsification is that greedy on a good reduced set $v$7 preserves the standard greedy approximation up to an additive error: $v$8 where $v$9 is the optimal size-$1-1/e$0 solution on the full ground set and $1-1/e$1 is the greedy solution on the reduced set (Zhou et al., 2016).

For randomized SS, the analysis states that under the sampling assumptions in Proposition 6, with high probability the output set $1-1/e$2 satisfies

$1-1/e$3

and greedy on $1-1/e$4 achieves

$1-1/e$5

with probability at least

$1-1/e$6

The degradation is therefore a factor $1-1/e$7 in the additive $1-1/e$8-term rather than a change in the multiplicative $1-1/e$9 form (Zhou et al., 2016).

The output size is also characterized. If $G(V,E,w)$0, where $G(V,E,w)$1 is the size of an optimal sparsified set for threshold $G(V,E,w)$2, then

$G(V,E,w)$3

Per iteration, the dominant cost is computing $G(V,E,w)$4 for all remaining $G(V,E,w)$5, which requires $G(V,E,w)$6 pairwise evaluations because $G(V,E,w)$7. The evaluations are described as highly parallelizable, and memory usage drops sharply because greedy is run only on $G(V,E,w)$8 (Zhou et al., 2016).

Application papers inherit this general profile but report domain-specific costs. The billboard-slot formulation gives total complexity

$G(V,E,w)$9

and space

$(u,v)$0

with the dense graph stored in an adjacency matrix (Ali et al., 2022). The RL formulation states that Algorithm 1: Submodularity Graph-based Policy Optimization (SGPO) requires

$(u,v)$1

time and

$(u,v)$2

space, with the final bound dominated by gradient estimation (Anand et al., 18 Jul 2025). The RL paper also proves an inapproximability result: guaranteeing

$(u,v)$3

in polynomial time is impossible for any constant $(u,v)$4, unless

$(u,v)$5

5. Domain-specific instantiations

The same graph-pruning pattern has been instantiated in summarization, influential billboard slot selection, and reinforcement learning.

Domain	Instantiation	Reported findings
News and video summarization (Zhou et al., 2016)	SS + greedy/lazy greedy	Relative utility $(u,v)$6; lower runtime and memory
Billboard slot selection (Ali et al., 2022)	PSG + preprocessing + greedy	Combined preprocessing + pruning removes about 99.7% to 99.95% of slots
Billboard slot selection with spatial clustering (Ali et al., 2023)	$(u,v)$7-partition + PSG + greedy	More influence than baselines within reasonable computational time
RL with submodular rewards (Anand et al., 18 Jul 2025)	SGPO	More reward than SubPO baseline; more stable training behavior

In summarization, the method was evaluated on NYT, DUC 2001, and 25 SumMe videos. On the NYT corpus, SS typically achieved relative utility $(u,v)$8, sometimes slightly exceeded the full greedy baseline, and grew much more slowly in runtime as $(u,v)$9 increased. On DUC 2001, SS matched lazy greedy almost exactly across multiple summary lengths and topics while outperforming sieve-streaming. On SumMe, SS consistently approached or outperformed lazy greedy in recall and F1 while significantly reducing runtime (Zhou et al., 2016).

In Influential Billboard Slot Selection (IBSS), the ground set is the set of billboard slots $w_{uv}=f(v|u)-f(u|V\backslash u).$0, and the objective is

$w_{uv}=f(v|u)-f(u|V\backslash u).$1

with influence

$w_{uv}=f(v|u)-f(u|V\backslash u).$2

The influence function is stated to be non-negative, monotone, and submodular, which justifies greedy after pruning. The 2022 paper applies zero-influence preprocessing, PSG pruning, and final incremental greedy, and reports that the proposed method generally achieves the highest influence, although Mall at $w_{uv}=f(v|u)-f(u|V\backslash u).$3 is noted as an exception where Top-$w_{uv}=f(v|u)-f(u|V\backslash u).$4 slightly outperforms (Ali et al., 2022).

The 2023 billboard paper adds a spatial clustering layer before PSG. It defines overlap

$w_{uv}=f(v|u)-f(u|V\backslash u).$5

and a $w_{uv}=f(v|u)-f(u|V\backslash u).$6-partition based on overlap ratio, then applies PSG and greedy on the merged surviving clusters. The paper reports that increasing $w_{uv}=f(v|u)-f(u|V\backslash u).$7 decreases influence quality but reduces time, and that the proposed methods consistently outperform all baselines in influence (Ali et al., 2023).

In RL, the ground object is a submodular MDP

$w_{uv}=f(v|u)-f(u|V\backslash u).$8

with expected return

$w_{uv}=f(v|u)-f(u|V\backslash u).$9

SGPO rolls out a trajectory, builds a submodularity graph over visited states, prunes low-divergence states, constructs a training set

$n$00

estimates the policy gradient using Theorem 2, and updates parameters using the regularized ascent rule in Eq. (4). With $n$01 and $n$02, the reported experiments show higher objective values than SubPO across Car Racing, MuJoCo Ant, graph-based, and entropy-based environments (Anand et al., 18 Jul 2025).

6. Distinctions, assumptions, and limitations

A recurrent source of confusion is the relation between pruned submodularity graphs and other graph-pruning paradigms. The neural method Prune4SED also performs graph pruning, but it does so by converting pruning into node relabeling, relaxing the discrete keep/prune variable to a continuous keep probability, and learning query-aware pruning end-to-end for subgraph edit distance. The paper explicitly states that it is not solving pruning via submodular optimization or classic combinatorial heuristics; instead it uses attention, hard/soft pruning, and multi-head prediction (Liu et al., 2022). The shared term “pruning” therefore does not imply shared optimization structure.

The pruned submodularity graph-based approach also depends on strong assumptions. Its guarantees rely on submodularity, and several application papers make this explicit: the billboard influence objective is proven non-negative, monotone, and submodular; the RL paper models the problem with a submodular reward functional; and the foundational analysis depends on sampling assumptions and the existence of a good sparsified set (Zhou et al., 2016). A plausible implication is that performance claims should not be transferred mechanically to objectives lacking diminishing returns.

Limitations are likewise domain dependent. Dense graph storage yields $n$03 space in both billboard and RL formulations (Ali et al., 2022). In RL, the broader problem is not constant-factor approximable under the stated complexity assumption, so pruning addresses feasible computation time rather than eliminating intrinsic hardness (Anand et al., 18 Jul 2025). In spatially clustered billboard selection, exact $n$04-partitioning is NP-hard and only approximate partitioning is used; the cluster pruning heuristic may discard useful clusters if $n$05 is too aggressive (Ali et al., 2023).

Across the literature, the unifying interpretation is consistent: the graph encodes pairwise substitutability or relative importance through submodular marginal gains; divergence measures how safely an element can be removed with respect to a retained set; and randomized or structured pruning reduces the effective problem size before the final optimization stage. Empirically, this often preserves near-greedy quality or improves downstream training behavior, but the formal guarantees remain additive-loss or approximate-preservation statements rather than exact equivalence (Zhou et al., 2016).