Balanced Signed Graph Denoisers

• Balanced signed graph denoisers are methods that enforce or recover structural balance in signed graphs by adjusting vertices, edges, or signal paths.
• They leverage spectral techniques, frustration index minimization, and negation set theory to correct imbalances and enhance graph integrity.
• Applications span robust graph cleaning, EEG signal processing, and balance-aware GNN propagation to mitigate oversmoothing and improve performance.

Balanced signed graph denoisers are methods that recover, enforce, or exploit structural balance in signed graphs, where edges carry positive or negative signs. In the classical Harary sense, a signed graph is balanced when its vertices can be partitioned into two groups such that all edges within groups are positive and all edges across groups are negative; equivalently, every cycle has positive sign product. In recent work, denoising is interpreted in several technically distinct but closely related ways: deleting vertices to obtain a balanced induced subgraph, flipping edge signs to reach a balanced state with low frustration, filtering signals on a balanced signed graph via its spectrum, or redesigning signed message passing so that attraction and repulsion respect a balanced organization (Ordozgoiti et al., 2020, Shebaro et al., 2023, Shebaro et al., 2023, Yao et al., 3 Oct 2025, Wang et al., 17 Feb 2025).

1. Structural balance and spectral foundations

A signed graph is variously written as $G=(V,E^+,E^-)$ or $\Sigma=(G,\sigma)$, with $\sigma:e\mapsto\{+1,-1\}$. Across the literature, the core balance criterion is consistent: a signed graph is balanced iff every cycle is positive, iff all paths between any vertex pair have the same sign, iff the frustration is zero, and iff there exists a Harary bipartition $U,W$ such that an edge is negative exactly when it goes between $U$ and $W$ (Shebaro et al., 2023, Shebaro et al., 2023). The EEG-denoising literature uses the equivalent formulation that a balanced signed graph has no cycle containing an odd number of negative edges and can be represented by node polarities $\beta_i\in\{1,-1\}$ satisfying

$$\beta_i\beta_j=\operatorname{sign}(w_{i,j}), \qquad \forall (i,j)\in E.$$

This polarity view is the basis for constructing balanced signed graph filters and for mapping signed Laplacians to positive-graph Laplacians (Yao et al., 3 Oct 2025).

Spectral formulations are central to balanced denoising. For a signed graph, the adjacency matrix is defined by

$$A_{ij}= \begin{cases} 1 & \{i,j\}\in E^+\ -1 & \{i,j\}\in E^-\ 0 & \text{otherwise}, \end{cases}$$

the degree matrix $D$ uses unsigned degrees, and the signed Laplacian is

$\Sigma=(G,\sigma)$0

A foundational result used by TIMBAL states that for a connected signed graph with signed Laplacian $\Sigma=(G,\sigma)$1 and eigenvalues $\Sigma=(G,\sigma)$2,

$\Sigma=(G,\sigma)$3

The same paper uses the bound

$\Sigma=(G,\sigma)$4

which operationalizes the idea that a small smallest eigenvalue indicates proximity to balance (Ordozgoiti et al., 2020).

For balanced signed graphs, the spectral structure can be transferred to an unsigned positive graph. Let

$\Sigma=(G,\sigma)$5

Then the balanced signed Laplacian $\Sigma=(G,\sigma)$6 is mapped to a positive graph Laplacian $\Sigma=(G,\sigma)$7 by

$\Sigma=(G,\sigma)$8

Because $\Sigma=(G,\sigma)$9, $\sigma:e\mapsto\{+1,-1\}$0 and $\sigma:e\mapsto\{+1,-1\}$1 have the same eigenvalues, and their eigenvectors satisfy $\sigma:e\mapsto\{+1,-1\}$2. This is the key spectral fact that makes ideal low-pass denoising well defined on balanced signed graphs (Yao et al., 3 Oct 2025).

2. Denoising objectives and measures of imbalance

The denoising target depends on what is treated as noise. One major formulation is the maximum balanced subgraph problem: given $\sigma:e\mapsto\{+1,-1\}$3, find an induced subgraph $\sigma:e\mapsto\{+1,-1\}$4 that is balanced and maximizes $\sigma:e\mapsto\{+1,-1\}$5,

$\sigma:e\mapsto\{+1,-1\}$6

In this formulation, the “noise” is the set of vertices whose presence prevents exact balance, and the minimum number of vertices that must be removed is the frustration number (Ordozgoiti et al., 2020). ABCD uses a closely related formulation in terms of a largest balanced subgraph $\sigma:e\mapsto\{+1,-1\}$7 satisfying $\sigma:e\mapsto\{+1,-1\}$8 while maximizing graph size (Shebaro et al., 2023).

A second formulation treats noise as incorrect edge signs. Here the central quantity is the frustration index, defined as the minimum number of candidate edge signs that must be switched for the graph to become balanced. In the tree-based and gradient-based recovery literature, denoising therefore means finding a nearby balanced graph by flipping as few edge signs as possible (Shebaro et al., 2023). The negation-set literature makes the same idea combinatorial: a negation set is a set of edges whose negation yields a balanced graph, and the negation sets of a signed graph are precisely the negative edge sets that can be obtained by switching the graph (Lacasse, 2020).

More recent GNN-oriented work introduces additional imbalance measures. Structural Balanced Propagation defines the structural imbalance degree

$\sigma:e\mapsto\{+1,-1\}$9

where $U,W$0 and $U,W$1 count same-label non-positive and different-label non-negative inconsistencies. A perfectly structurally balanced graph has $U,W$2 (Wang et al., 17 Feb 2025). SiGAug uses an edge-local utility notion for structural denoising. For a negative edge $U,W$3, if $U,W$4 is the set of all cycles containing $U,W$5 of length at most $U,W$6 and $U,W$7 its balanced subset, then

$U,W$8

and in practice

$U,W$9

Edges below a threshold $U$0 are removed (Chen et al., 2023).

These formulations imply different denoising semantics. Vertex-deletion methods produce a balanced induced subgraph, edge-flip methods produce a balanced state on the same graph, cycle-utility filters remove structurally harmful negatives, and signal filters assume a balanced signed graph is already available or is constructed by design. This suggests that “balanced signed graph denoiser” is best understood as a family of operators organized by what object is corrected: vertices, edges, signals, or propagation dynamics.

3. Vertex-deletion denoisers for maximum balanced subgraphs

TIMBAL is an efficient algorithm for finding large balanced induced subgraphs by greedily trimming vertices that most reduce spectral imbalance (Ordozgoiti et al., 2020). Its ideal criterion is

$U$1

where $U$2 is the Laplacian after deleting vertex $U$3. To avoid solving a new eigenproblem for every candidate deletion, TIMBAL derives the perturbation bound

$U$4

where $U$5 is the unit eigenvector associated with $U$6. The corresponding ranking vector can be computed as

$U$7

with $U$8 the entrywise absolute value of $U$9, $W$0, and $W$1. The algorithm also allows batch removal of pairwise nonadjacent vertices via an additive-type bound, then performs a restoration stage: once a balanced induced subgraph and a consistent partition $W$2 are found, removed vertices are revisited in deletion order and reintroduced whenever they can be assigned to one side without violating balance. The paper frames this as removing structural outliers or inconsistent samples and then adding back compatible ones. For scalability, TIMBAL uses randomized BFS sampling of small connected subgraphs, runs the trimming method on each, and unions the vertices identified for removal. It is tested on real signed networks and on enlarged graphs up to about $W$3 million vertices and $W$4 million edges, with runtimes on the order of minutes to under two hours.

ABCD, “Algorithm for Balanced Component Discovery,” also targets the largest balanced subgraph but does so through a two-stage procedure that searches over a small set of near-balanced states and then purges vertices while preserving connectivity (Shebaro et al., 2023). In Phase 1, it samples $W$5 spanning trees, examines all non-tree edges, and records the candidate edge sets $W$6 whose flips would balance the corresponding fundamental cycles. It then keeps only the top $W$7 states with smallest frustration, storing $W$8 for frustrations, $W$9 for edge sets to flip, and $\beta_i\in\{1,-1\}$0 for the Harary bit-vectors. In Phase 2, it constructs vertex deletion sets $\beta_i\in\{1,-1\}$1 so that

$\beta_i\in\{1,-1\}$2

and then returns the largest connected component $\beta_i\in\{1,-1\}$3 among the $\beta_i\in\{1,-1\}$4 candidates. Three variants are defined: ABCDD, which deletes the endpoint with smaller degree; ABCDH, which uses the Harary bipartition and degree/neighborhood-degree heuristics; and ABCDS, which uses an adjusted status

$\beta_i\in\{1,-1\}$5

Its reported complexity is

$\beta_i\in\{1,-1\}$6

with additional terms depending on the variant. Empirically, ABCD matches TIMBAL on Highland and ProLeague, finds a larger subgraph on 13 KONECT datasets, and loses only on WikiElec in that benchmark. On PPI, TIMBAL returns 900 vertices, while ABCDH, ABCDS, and ABCDD return 2,073, 2,038, and 2,149 respectively; on Epinions, TIMBAL returns 73,433 (5 runs) or 74,106 (10 runs), while ABCDS returns 78,126. On Amazon graphs, TIMBAL is often unable to finish within 48 hours, whereas ABCD completes and returns much larger balanced subgraphs, including 7,458,256 vertices on Books and 3,689,985 on Electronics with ABCDD.

Both TIMBAL and ABCD instantiate denoising by deletion rather than sign correction. They do not relax balance; instead, they remove the smallest disruptive portion of the graph that prevents a zero-frustration, exactly balanced remainder.

4. Edge-flip correction and negation-set methods

A different line of work denoises signed graphs by editing edge signs. “Scaling Frustration Index and Corresponding Balanced State Discovery for Real Signed Graphs” formalizes this as minimizing the frustration index and proposes two scalable approximations: the tree-based graphBpp and the gradient-descent-based graphL (Shebaro et al., 2023). graphBpp samples spanning trees, forms the fundamental cycle $\beta_i\in\{1,-1\}$7 for each non-tree edge $\beta_i\in\{1,-1\}$8, flips $\beta_i\in\{1,-1\}$9 if the cycle is negative, counts the flips, and stores the best balanced states in a memory-bounded “frustration cloud”

$$\beta_i\beta_j=\operatorname{sign}(w_{i,j}), \qquad \forall (i,j)\in E.$$0

Because a connected component has exactly

$$\beta_i\beta_j=\operatorname{sign}(w_{i,j}), \qquad \forall (i,j)\in E.$$1

fundamental cycles, the method avoids exhaustive cycle enumeration. Its complexity is given as

$$\beta_i\beta_j=\operatorname{sign}(w_{i,j}), \qquad \forall (i,j)\in E.$$2

it uses at most $$\beta_i\beta_j=\operatorname{sign}(w_{i,j}), \qquad \forall (i,j)\in E.$$3 of total RAM in the reported experiments, and it compares seven tree samplers: BFS, DFS, RDFS, Aldous-Broder, Kruskal, Prim, and Hybrid RDFS-BFS. The paper concludes that BFS is generally the best practical default. graphL instead minimizes the relaxed imbalance loss

$$\beta_i\beta_j=\operatorname{sign}(w_{i,j}), \qquad \forall (i,j)\in E.$$4

with gradient

$$\beta_i\beta_j=\operatorname{sign}(w_{i,j}), \qquad \forall (i,j)\in E.$$5

followed by thresholding $$\beta_i\beta_j=\operatorname{sign}(w_{i,j}), \qquad \forall (i,j)\in E.$$6 to $$\beta_i\beta_j=\operatorname{sign}(w_{i,j}), \qquad \forall (i,j)\in E.$$7 and flipping edges that violate the inferred bipartition. Its per-iteration cost is

$$\beta_i\beta_j=\operatorname{sign}(w_{i,j}), \qquad \forall (i,j)\in E.$$8

and it uses CSR sparse storage rather than an adjacency matrix. The paper reports speedups around 300+ times faster than the state-of-the-art on large graphs; graphBpp processes Amazon Books, with about 10 million vertices and over 22 million edges, in 19 hours, while graphL is much faster and often finds lower frustration indices, albeit with sensitivity to initialization and hyperparameters.

The negation-set literature provides a more structural theory of edge correction (Lacasse, 2020). A negation set is minimal if no proper subset is also a negation set, minimum if it has minimum cardinality, and uniquely minimum if no other negation set has the same minimum size. A sharp test for minimality is

$$\beta_i\beta_j=\operatorname{sign}(w_{i,j}), \qquad \forall (i,j)\in E.$$9

The paper also proves that a negation set is bipartite iff there exists another negation set disjoint from it, establishes sufficient conditions for minimum and unique minimum negation sets, and gives an algorithm for computing a maximum family of pairwise disjoint negation sets containing a given one. For a bipartite negation set $$A_{ij}= \begin{cases} 1 & \{i,j\}\in E^+\ -1 & \{i,j\}\in E^-\ 0 & \text{otherwise}, \end{cases}$$0, if $$A_{ij}= \begin{cases} 1 & \{i,j\}\in E^+\ -1 & \{i,j\}\in E^-\ 0 & \text{otherwise}, \end{cases}$$1 is the set of stable bipartitions of $$A_{ij}= \begin{cases} 1 & \{i,j\}\in E^+\ -1 & \{i,j\}\in E^-\ 0 & \text{otherwise}, \end{cases}$$2, then

$$A_{ij}= \begin{cases} 1 & \{i,j\}\in E^+\ -1 & \{i,j\}\in E^-\ 0 & \text{otherwise}, \end{cases}$$3

where $$A_{ij}= \begin{cases} 1 & \{i,j\}\in E^+\ -1 & \{i,j\}\in E^-\ 0 & \text{otherwise}, \end{cases}$$4 is the minimum positive distance between the two classes. This theory does not present a generic large-scale denoiser by itself, but it characterizes when edge-negation corrections are minimal, unique, or packable, and therefore supplies exact combinatorial constraints for sign-flip denoising.

5. Signal denoisers on balanced signed graphs

Balanced signed graph denoisers also appear as signal-processing operators rather than graph-editing procedures. “Lightweight Transformer for EEG Classification via Balanced Signed Graph Algorithm Unrolling” constructs balanced signed graphs from EEG sensor relations, maps them to positive graphs through a similarity transform, and applies spectral low-pass denoising to the signals (Yao et al., 3 Oct 2025). The graph is built from learned feature distances

$$A_{ij}= \begin{cases} 1 & \{i,j\}\in E^+\ -1 & \{i,j\}\in E^-\ 0 & \text{otherwise}, \end{cases}$$5

with signed weights chosen so that the sign agrees with the polarity relation and therefore enforces balance: $$A_{ij}= \begin{cases} 1 & \{i,j\}\in E^+\ -1 & \{i,j\}\in E^-\ 0 & \text{otherwise}, \end{cases}$$6 To ensure positive semidefiniteness, the method adds a diagonal shift based on Gershgorin bounds,

$$A_{ij}= \begin{cases} 1 & \{i,j\}\in E^+\ -1 & \{i,j\}\in E^-\ 0 & \text{otherwise}, \end{cases}$$7

The denoising objective is projection onto the low-frequency subspace of the corresponding positive graph: $$A_{ij}= \begin{cases} 1 & \{i,j\}\in E^+\ -1 & \{i,j\}\in E^-\ 0 & \text{otherwise}, \end{cases}$$8 whose solution is

$$A_{ij}= \begin{cases} 1 & \{i,j\}\in E^+\ -1 & \{i,j\}\in E^-\ 0 & \text{otherwise}, \end{cases}$$9

with ideal low-pass response

$D$0

To avoid full eigendecomposition, the paper uses Lanczos approximation: $D$1 The hard spectral cutoff is replaced during training by a differentiable sigmoid

$D$2

so each low-pass block learns its own cutoff $D$3.

Architecturally, the network alternates a Balanced Graph Learning (BGL) module with a Low-Pass Filter (LPF) module. BGL extracts node features with a shallow CNN, computes pairwise distances via a learnable metric matrix $D$4, converts distances to signed weights, updates polarities to preserve balance, and maps $D$5 to $D$6. LPF then performs the learned ideal low-pass filtering via Lanczos approximation. The model trains two class-specific denoisers,

$D$7

and classifies by reconstruction error,

$D$8

The paper reports 97.57% accuracy on the main EEG task with only 14,787 parameters, versus a transformer baseline with over 1.8M parameters and lower accuracy, as well as 90.06% accuracy and 92.59% F1 under LOSO evaluation. It also reports that balanced signed graphs outperform both positive graphs and unbalanced signed graphs.

In this setting, denoising is neither vertex deletion nor sign correction. Instead, balance is enforced at graph construction time so that signal frequencies are well defined, after which denoising becomes an orthogonal low-pass projection on the associated positive graph.

6. Balance-aware signed GNN denoising and propagation

The signed-GNN literature reinterprets denoising as the control of attractive and repulsive propagation. “A Signed Graph Approach to Understanding and Mitigating Oversmoothing in GNNs” studies the linear propagation rule

$D$9

or nodewise

$\Sigma=(G,\sigma)$00

Positive edges average features and negative edges subtract neighbors’ messages, creating repulsion. The paper proves that if repulsion is too weak, the system converges to average consensus and oversmoothing; if it is too strong, the norm diverges for almost all initializations. It then introduces Structural Balanced Propagation (SBP), which explicitly constructs signed graphs approximating structural balance. For a structurally balanced complete graph with bounded activation and sufficiently large $\Sigma=(G,\sigma)$01, Theorem 4.3 states that one balance class converges to $\Sigma=(G,\sigma)$02 and the other to $\Sigma=(G,\sigma)$03 with probability one. Label-SBP and Feature-SBP construct the negative graph either from labels or from feature similarity, and experiments on nine benchmarks show that SBP improves classification accuracy and mitigates oversmoothing even at depths of up to 300 layers (Wang et al., 17 Feb 2025).

SiGAug addresses denoising through augmentation and cycle-aware filtering for link sign prediction (Chen et al., 2023). It separates two problems: semantic imbalance, because negative edges are much rarer than positive ones, and structural unbalance, because unbalanced triangles and quadrilaterals create contradictory local configurations. The method has three components: an Edge Utility Filter (EUF) that keeps perturbed negative edges with high utility, an Edge Perturbation Regulator (EPR) that balances the number of perturbed positive and negative edges and controls the perturbation rate, and downstream SGNN training on the augmented graph. The perturbation ratios are

$\Sigma=(G,\sigma)$04

With $\Sigma=(G,\sigma)$05, only triangles and quadrilaterals are counted for utility, and the reported default values are

$\Sigma=(G,\sigma)$06

Across Bitcoin_OTC, Bitcoin_Alpha, Congress, Chess, and Wiki_Election, SiGAug improves SGNN backbones such as SGCN, RSGNN, and SNEA. With SGCN, the reported AUC improvement over the original backbone is about 16% to 72%; overall F1 gains are about 11% to 44%, and negative-edge F1 gains are about 29% to 264%.

RIDGE, by contrast, explicitly questions the idea that robust signed graph denoising should primarily mean making a graph more balanced (Wu et al., 26 Oct 2025). It studies link sign prediction on noisy signed graphs $\Sigma=(G,\sigma)$07 and extends Graph Information Bottleneck to joint input-target denoising. The proposed GIB with target-space denoising (GIB-TD) objective is

$\Sigma=(G,\sigma)$08

derived from the upper bound

$\Sigma=(G,\sigma)$09

RIDGE uses feature masking

$\Sigma=(G,\sigma)$10

a parameterized sampler $\Sigma=(G,\sigma)$11 to compute edge probabilities

$\Sigma=(G,\sigma)$12

and Bernoulli sampling to obtain cleaned topology and labels,

$\Sigma=(G,\sigma)$13

The final objective is

$\Sigma=(G,\sigma)$14

On Bitcoin_OTC, Bitcoin_Alpha, Epinions, and Slashdot, with noise ratios of 0%, 10%, 20%, and 25%, RIDGE reports up to 5.45% improvement over RSGNN and up to 4.95% improvement over LRWSB. Importantly, the paper states that RIDGE does not necessarily increase balance degree, yet still improves performance, and that the datasets contain many unbalanced triangles.

7. Limits, misconceptions, and open problems

The literature imposes substantial assumptions. TIMBAL targets exact balance in an induced subgraph rather than an approximately balanced objective, assumes an undirected simple graph, relies on the smallest Laplacian eigenpair, and treats optimal simultaneous batch removal as an open hard problem; disconnected graphs are handled by keeping the largest connected component or recursing (Ordozgoiti et al., 2020). ABCD is likewise not an exact global solver for the NP-hard maximum balanced subgraph problem; it searches only the top $\Sigma=(G,\sigma)$15 balanced states with lowest frustration and then applies heuristic vertex-removal criteria (Shebaro et al., 2023). graphBpp depends on spanning-tree sampling, while graphL requires appropriate initialization and hyperparameter tuning and returns only one balanced state rather than a cloud of alternatives (Shebaro et al., 2023).

Two recurring simplifications are explicitly contradicted by recent work. The first is that merely adding negative edges is enough to obtain useful repulsion. SBP shows that if $\Sigma=(G,\sigma)$16, propagation still collapses to average consensus, whereas if $\Sigma=(G,\sigma)$17, it diverges for almost all initializations; only when positive and negative edges are organized in a structurally balanced way does the dynamics preserve cluster separation (Wang et al., 17 Feb 2025). The second is that robustness is equivalent to maximizing balance. RIDGE reports that performance improves even though balance degree does not necessarily increase, and attributes this to the prevalence of unbalanced triangles and noisy supervision in real signed graphs (Wu et al., 26 Oct 2025).

A further unresolved issue is the relation between exact balance and practical utility. TIMBAL notes that intermediate iterates can be used when some imbalance is tolerable, and ABCD shows that increasing $\Sigma=(G,\sigma)$18 improves subgraph size at the cost of runtime, indicating a balance–compute tradeoff rather than a single canonical optimum (Ordozgoiti et al., 2020, Shebaro et al., 2023). In signal denoising, the balanced signed graph must often be constructed rather than recovered, as in the EEG setting where polarities and edge signs are chosen to satisfy balance by design (Yao et al., 3 Oct 2025). In signed GNNs, SiGAug, SBP, and RIDGE together suggest that denoising may mean selective removal of structurally harmful negative edges, explicit balancing of attractive and repulsive relations, or compression of task-irrelevant noise in both graph and target spaces, depending on the task and the assumed noise model (Chen et al., 2023, Wang et al., 17 Feb 2025, Wu et al., 26 Oct 2025).

Taken together, these results define balanced signed graph denoisers not as a single algorithmic class but as a spectrum of methods centered on structural balance: exact balanced-subgraph extraction, frustration-minimizing state recovery, sign-correction via negation sets, balanced spectral filtering of graph signals, and balance-aware or balance-critical propagation in signed neural architectures. The common principle is that imbalance is treated as a structured defect rather than generic random noise, but the literature also makes clear that the appropriate correction operator depends on whether the goal is graph cleaning, state recovery, signal denoising, or robust prediction.