Edge Score: Definitions & Applications

• Edge Score is a family of quantitative metrics that assign values to graph edges, image pixels, or transitions for comparison and decision support.
• It is employed in diverse settings such as dynamic graph analysis, connectivity significance, and spectral intervention using techniques like normalized Jaccard and PageRank.
• Applications span network theory, computer vision edge detection, and neural routing, each with domain-specific scoring criteria and performance metrics.

Searching arXiv for recent and foundational papers relevant to the multiple meanings of “Edge Score.” Edge Score denotes a family of quantitative constructions in which an edge, an edge set, a directed transition, or an edge pixel is assigned a value for comparison, ranking, prediction, or optimization. In graph dynamics, the score can be a normalized turnover of edge sets across time; in network science, it can be a significance, centrality, or intervention value of a single edge; in computer vision, it can be a dense per-pixel edge confidence; and in neural routing, it can be a transition-aware decoder logit rather than a node-only compatibility score (Bridonneau et al., 2023, Zhao et al., 2019, Kucharczuk et al., 2021, Li et al., 2023, Xie et al., 2015, Liang et al., 1 Jun 2026). The term therefore has no single domain-independent definition. Its meaning is determined by what object is being scored, what information is available, and what downstream decision the score is meant to support.

1. Semantic range and formal scope

Across the cited literature, “edge score” ranges from a scalar attached to an individual graph edge to a value attached to an entire edge set, to a dense confidence defined over image pixels that are hypothesized to lie on edges. In graph-theoretic settings, the scored object is typically either a single edge $e_{ij}$ or an edge set $E_t$. In vision, the scored object is usually a pixel in an edge map. In asymmetric routing, the scored object is not a node in isolation but the directed action “go from the current node to $j$” under the current partial route (Xie et al., 2015, Liang et al., 1 Jun 2026).

Setting	Scored object	Representative definition
Dynamic graphs	Edge-set change across snapshots	$$\mathcal{D}^e_t=\frac{|E_{t+1}\triangle E_t|}{|E_{t+1}\cup E_t|}$$
Connectivity significance	Single graph edge	$$SN(i,j)=\frac{|n^{(2)}_{i\backslash j}\cap n^{(2)}_{j\backslash i}|}{|n^{(2)}_{i\backslash j}\cup n^{(2)}_{j\backslash i}|}$$
Recursive centrality	Directed edge	$$PR^a_e(G)=\frac{1}{\deg_u^+(G)}\left(a\sum_{e'\in E^-_u(G)}PR^a_{e'}(G)+b(u)\right)$$
Spectral intervention	Weighted edge entry	$\mathrm{EC}_r(i,j;X)=(\tilde X_r)_{ij}$
Vision and routing	Pixel or directed transition	$\hat Y_{\text{fuse}}$ or $\tilde s_t(j)=s_t^{\mathrm{base}}(j)+b_t(j)$

Several works explicitly delimit what does not count as a direct edge score. The ordering-based causal discovery framework via generalized score matching does not define a per-edge score $s(i\to j)$; it computes node-level leaf discriminants and then uses the recovered order to constrain edge recovery (Vo et al., 22 Jan 2026). In HMGAT for hybrid beamforming, the closest edge-related quantities are attention compatibility scores $E_t$0 and $E_t$1, while the separate phrase “score-based” refers to gradients of log-densities in CSI generation and denoising, not to graph-edge scoring (Li et al., 10 Nov 2025). In Acc3D, “edge consistency” refers to a high-SNR region near a diffusion endpoint rather than to image or graph edges (Liu et al., 20 Mar 2025).

2. Edge-set turnover in dynamic graphs

For dynamic graphs represented as a sequence of snapshots $E_t$2, the edge-level DynamicScore, or E-DynamicScore, is defined as

$E_t$3

with $E_t$4 (Bridonneau et al., 2023). The required inputs are only the two consecutive edge sets $E_t$5 and $E_t$6. The numerator counts edges present in exactly one snapshot, and the denominator counts all edges present in at least one of the two snapshots. The score is bounded in $E_t$7: $E_t$8 means identical edge sets and $E_t$9 means no common edges.

This definition is explicitly Jaccard-distance-like and is intended to capture both changes in edge-set cardinality and changes in edge-set composition. That distinction is central. A simple count-based quantity such as $j$0 can be zero even when every old edge disappears and is replaced by a new one, whereas $j$1 remains large in exactly that case (Bridonneau et al., 2023).

In the Barabási–Albert preferential attachment model considered in the paper, one new vertex is added at each step and connects to $j$2 existing vertices. Under that setup,

$j$3

and the paper concludes that the score decreases and tends toward $j$4 as $j$5 (Bridonneau et al., 2023). The interpretation given is that the graph keeps growing, but the relative amount of edge change per step becomes smaller and smaller.

For Edge-Markovian graphs, each possible edge independently follows a two-state Markov chain with transition matrix

$j$6

The proof in the paper gives the average symmetric-difference size as

$j$7

and the average union size as

$j$8

so the average E-DynamicScore is the ratio of those quantities (Bridonneau et al., 2023). At stationary density $j$9, the paper gives

$$\mathcal{D}^e_t=\frac{|E_{t+1}\triangle E_t|}{|E_{t+1}\cup E_t|}$$0

states that it can take all values from $$\mathcal{D}^e_t=\frac{|E_{t+1}\triangle E_t|}{|E_{t+1}\cup E_t|}$$1 to $$\mathcal{D}^e_t=\frac{|E_{t+1}\triangle E_t|}{|E_{t+1}\cup E_t|}$$2, is independent of $$\mathcal{D}^e_t=\frac{|E_{t+1}\triangle E_t|}{|E_{t+1}\cup E_t|}$$3, and decreases as $$\mathcal{D}^e_t=\frac{|E_{t+1}\triangle E_t|}{|E_{t+1}\cup E_t|}$$4 increases (Bridonneau et al., 2023). The reported simulations support that statement: for fixed $$\mathcal{D}^e_t=\frac{|E_{t+1}\triangle E_t|}{|E_{t+1}\cup E_t|}$$5, varying $$\mathcal{D}^e_t=\frac{|E_{t+1}\triangle E_t|}{|E_{t+1}\cup E_t|}$$6 leaves the average E-DynamicScore essentially unchanged.

3. Connectivity significance and recursive centrality

One major use of edge scores is to identify edges that are structurally important for connectivity. The second-order neighborhood index $$\mathcal{D}^e_t=\frac{|E_{t+1}\triangle E_t|}{|E_{t+1}\cup E_t|}$$7 is defined for connected, simple, unweighted, undirected graphs by first deleting the target edge $$\mathcal{D}^e_t=\frac{|E_{t+1}\triangle E_t|}{|E_{t+1}\cup E_t|}$$8, then computing the sets of nodes at distance exactly $$\mathcal{D}^e_t=\frac{|E_{t+1}\triangle E_t|}{|E_{t+1}\cup E_t|}$$9 from each endpoint, and finally taking their Jaccard similarity: $$SN(i,j)=\frac{|n^{(2)}_{i\backslash j}\cap n^{(2)}_{j\backslash i}|}{|n^{(2)}_{i\backslash j}\cup n^{(2)}_{j\backslash i}|}$$0 The paper’s interpretation is inverted relative to many centralities: smaller $$SN(i,j)=\frac{|n^{(2)}_{i\backslash j}\cap n^{(2)}_{j\backslash i}|}{|n^{(2)}_{i\backslash j}\cup n^{(2)}_{j\backslash i}|}$$1 means a more significant edge, because weak overlap of second-order neighborhoods suggests that the edge bridges different communities or subnetworks (Zhao et al., 2019). Evaluated by targeted edge percolation on 15 real networks, $$SN(i,j)=\frac{|n^{(2)}_{i\backslash j}\cap n^{(2)}_{j\backslash i}|}{|n^{(2)}_{i\backslash j}\cup n^{(2)}_{j\backslash i}|}$$2 yields the lowest robustness $$SN(i,j)=\frac{|n^{(2)}_{i\backslash j}\cap n^{(2)}_{j\backslash i}|}{|n^{(2)}_{i\backslash j}\cup n^{(2)}_{j\backslash i}|}$$3 on all 15, outperforming edge betweenness, degree product, diffusion intensity, topological overlap, and bridgeness (Zhao et al., 2019).

A different conception is recursive feedback centrality. Edge PageRank is defined on directed multigraphs with node weights and possible self-loops. For an edge $$SN(i,j)=\frac{|n^{(2)}_{i\backslash j}\cap n^{(2)}_{j\backslash i}|}{|n^{(2)}_{i\backslash j}\cup n^{(2)}_{j\backslash i}|}$$4,

$$SN(i,j)=\frac{|n^{(2)}_{i\backslash j}\cap n^{(2)}_{j\backslash i}|}{|n^{(2)}_{i\backslash j}\cup n^{(2)}_{j\backslash i}|}$$5

with $$SN(i,j)=\frac{|n^{(2)}_{i\backslash j}\cap n^{(2)}_{j\backslash i}|}{|n^{(2)}_{i\backslash j}\cup n^{(2)}_{j\backslash i}|}$$6 (Kucharczuk et al., 2021). The score of an edge depends on PageRank mass arriving into its start node, the node’s baseline weight, and division by the start node’s out-degree. The paper shows that an edge centrality measure satisfies Node Deletion, Edge Deletion, Edge Multiplication, Edge Swap, Node Redirect, and Baseline if and only if it is Edge PageRank (Kucharczuk et al., 2021). This makes Edge PageRank the unique score under that axiomatic package.

ECHO defines edge centrality as the optimizer of a neighborhood-based objective that combines a degree-based prior with adjacent-edge smoothness. Its closed form is

$$SN(i,j)=\frac{|n^{(2)}_{i\backslash j}\cap n^{(2)}_{j\backslash i}|}{|n^{(2)}_{i\backslash j}\cup n^{(2)}_{j\backslash i}|}$$7

where

$$SN(i,j)=\frac{|n^{(2)}_{i\backslash j}\cap n^{(2)}_{j\backslash i}|}{|n^{(2)}_{i\backslash j}\cup n^{(2)}_{j\backslash i}|}$$8

The score is therefore high when an edge joins low-outdegree endpoints and is reinforced by neighboring edges through $$SN(i,j)=\frac{|n^{(2)}_{i\backslash j}\cap n^{(2)}_{j\backslash i}|}{|n^{(2)}_{i\backslash j}\cup n^{(2)}_{j\backslash i}|}$$9 (Yang, 2024). The paper provides three $$PR^a_e(G)=\frac{1}{\deg_u^+(G)}\left(a\sum_{e'\in E^-_u(G)}PR^a_{e'}(G)+b(u)\right)$$0-per-iteration estimation algorithms, additive per-edge error guarantees for the two iterative summation methods, and experiments showing that ECHO is the fastest method on all six tested datasets while offering superior practical effectiveness on most downstream tasks (Yang, 2024).

4. Spectral intervention scores on weighted and time-varying networks

In weighted networks, edge score can be defined as a first-order marginal value for reducing spectral objectives. For a weighted, possibly directed graph with nonnegative weight matrix $$PR^a_e(G)=\frac{1}{\deg_u^+(G)}\left(a\sum_{e'\in E^-_u(G)}PR^a_{e'}(G)+b(u)\right)$$1, the spectral objective is

$$PR^a_e(G)=\frac{1}{\deg_u^+(G)}\left(a\sum_{e'\in E^-_u(G)}PR^a_{e'}(G)+b(u)\right)$$2

where $$PR^a_e(G)=\frac{1}{\deg_u^+(G)}\left(a\sum_{e'\in E^-_u(G)}PR^a_{e'}(G)+b(u)\right)$$3 is the $$PR^a_e(G)=\frac{1}{\deg_u^+(G)}\left(a\sum_{e'\in E^-_u(G)}PR^a_{e'}(G)+b(u)\right)$$4-th largest singular value of $$PR^a_e(G)=\frac{1}{\deg_u^+(G)}\left(a\sum_{e'\in E^-_u(G)}PR^a_{e'}(G)+b(u)\right)$$5 (Li et al., 2023). The generalized edge centrality score at matrix $$PR^a_e(G)=\frac{1}{\deg_u^+(G)}\left(a\sum_{e'\in E^-_u(G)}PR^a_{e'}(G)+b(u)\right)$$6 is

$$PR^a_e(G)=\frac{1}{\deg_u^+(G)}\left(a\sum_{e'\in E^-_u(G)}PR^a_{e'}(G)+b(u)\right)$$7

and the gradient entry is

$$PR^a_e(G)=\frac{1}{\deg_u^+(G)}\left(a\sum_{e'\in E^-_u(G)}PR^a_{e'}(G)+b(u)\right)$$8

When $$PR^a_e(G)=\frac{1}{\deg_u^+(G)}\left(a\sum_{e'\in E^-_u(G)}PR^a_{e'}(G)+b(u)\right)$$9, this reduces to the classical product of the two endpoint eigenscores or singular-vector scores.

This score underlies a convex budgeted intervention problem in which existing edge weights are reduced under box constraints and a total budget $\mathrm{EC}_r(i,j;X)=(\tilde X_r)_{ij}$0. The Frank-Wolfe direction subproblem is solved by the paper’s Top-K-EdgeCentrality rule: allocate the reduction budget to the highest-scoring edges first, fully reducing some and partially reducing the last one if necessary (Li et al., 2023). The iterative algorithm, called Frank-Wolfe-EC, repeatedly recomputes the truncated rank-$\mathrm{EC}_r(i,j;X)=(\tilde X_r)_{ij}$1 SVD of the current matrix, rescoring edges after each update.

The paper also extends the construction to time-varying graphs with matrices $\mathrm{EC}_r(i,j;X)=(\tilde X_r)_{ij}$2, where the objective depends on the product $\mathrm{EC}_r(i,j;X)=(\tilde X_r)_{ij}$3. The score matrix for time $\mathrm{EC}_r(i,j;X)=(\tilde X_r)_{ij}$4 becomes

$\mathrm{EC}_r(i,j;X)=(\tilde X_r)_{ij}$5

so an edge at time $\mathrm{EC}_r(i,j;X)=(\tilde X_r)_{ij}$6 is important to the extent that it contributes, through preceding and succeeding temporal matrices, to the top $\mathrm{EC}_r(i,j;X)=(\tilde X_r)_{ij}$7 singular modes of the overall product dynamics (Li et al., 2023).

Empirically, the paper reports that on eleven weighted static networks the method reduces infections by an average of $\mathrm{EC}_r(i,j;X)=(\tilde X_r)_{ij}$8 more than baselines and reduces the largest singular value by $\mathrm{EC}_r(i,j;X)=(\tilde X_r)_{ij}$9 more on average; the best choice of $\hat Y_{\text{fuse}}$0 outperforms $\hat Y_{\text{fuse}}$1 by $\hat Y_{\text{fuse}}$2 on average; and iterative recomputation outperforms the one-step variant by $\hat Y_{\text{fuse}}$3 on infections on average (Li et al., 2023). This makes edge score not merely a ranking heuristic, but an explicit spectral sensitivity.

5. Dense edge confidence, perceptual edge quality, and transition logits

In computer vision, edge score usually denotes a dense per-pixel confidence rather than a graph-theoretic quantity. Holistically-Nested Edge Detection defines side-output edge scores as sigmoid-transformed activations

$\hat Y_{\text{fuse}}$4

and a fused edge score

$\hat Y_{\text{fuse}}$5

These are continuous confidence maps over pixels, not hard decisions. The paper reports ODS F-score .782 on BSD500 and .746 on NYU Depth, with about 0.4 second per image (Xie et al., 2015).

RankED reinterprets the predicted scalar at each pixel as an edge score to be ranked. It uses an AP-style ranking loss that pushes positive pixels above negatives and a sorting loss that pushes high-certainty edge pixels above lower-certainty ones: $\hat Y_{\text{fuse}}$6 The certainty map $\hat Y_{\text{fuse}}$7 is derived from multiple annotators with a spatial tolerance, and the paper explicitly does not threshold away uncertain positives during training, using $\hat Y_{\text{fuse}}$8 instead (Cetinkaya et al., 2024). On BSDS500, RankED reports AP .895 in the single-scale setting and .911 in the multi-scale setting; on Multi-cue edge labels, RankED reports AP .973 (Cetinkaya et al., 2024). The output is therefore best read as a certainty-aware ranking score rather than a calibrated posterior probability.

A different vision use is reference-based edge-map quality assessment. The JND-based edge map measure, JNDBEM, uses a psychophysically estimated Just-Noticeable-Difference of $\hat Y_{\text{fuse}}$9 pixels so that displacements below that threshold are treated as perceptually negligible: $\tilde s_t(j)=s_t^{\mathrm{base}}(j)+b_t(j)$0 with $\tilde s_t(j)=s_t^{\mathrm{base}}(j)+b_t(j)$1 and $\tilde s_t(j)=s_t^{\mathrm{base}}(j)+b_t(j)$2 if the minimum displacement is below the JND (Ahmad et al., 2022). Here the edge score is a quality score for an entire candidate edge map relative to ground truth.

SuperEdge again returns per-pixel confidence maps, now in a self-supervised setting. It separates prediction into pixel-level and object-level edge branches, fuses them, and evaluates the resulting confidence maps with ODS, OIS, and AP (Kai et al., 2024). On BIPEDv2, the paper states improvements of $\tilde s_t(j)=s_t^{\mathrm{base}}(j)+b_t(j)$3 in ODS and $\tilde s_t(j)=s_t^{\mathrm{base}}(j)+b_t(j)$4 in OIS over STEdge (Kai et al., 2024).

In neural asymmetric routing, the score shifts from pixels to actions. The final candidate score is an additive transition-aware logit

$\tilde s_t(j)=s_t^{\mathrm{base}}(j)+b_t(j)$5

where the bias $\tilde s_t(j)=s_t^{\mathrm{base}}(j)+b_t(j)$6 is computed from explicit transition features including the current directed edge, the reverse edge, an antisymmetry contrast, a closure term, and static lookahead (Liang et al., 1 Jun 2026). The paper’s strongest ablation evidence is that removing the current-edge terms makes performance worse than the baseline, which supports the claim that the crucial signal is decision-time exposure of the current directed edge (Liang et al., 1 Jun 2026).

6. Edge-aware representation learning and boundary cases

In edge-attributed bipartite graphs, EAGLE learns an edge embedding $\tilde s_t(j)=s_t^{\mathrm{base}}(j)+b_t(j)$7 rather than a predeclared scalar centrality. Its theoretical optimum has the form

$\tilde s_t(j)=s_t^{\mathrm{base}}(j)+b_t(j)$8

where $\tilde s_t(j)=s_t^{\mathrm{base}}(j)+b_t(j)$9 is an edge-wise transition matrix induced by shared endpoints and $s(i\to j)$0 is a transformed edge-attribute matrix (Wang et al., 2024). The practical factorized feature propagation scheme computes $s(i\to j)$1, and the dual-view version learns separate $s(i\to j)$2- and $s(i\to j)$3-side propagations before combining them. For semi-supervised edge classification, the model outputs

$s(i\to j)$4

so the most natural edge score is the class probability assigned to a target label (Wang et al., 2024).

The Edged Weisfeiler–Lehman algorithm and its neural counterparts EGIN, EGIN-C, and EGIN-E do not define a scalar edge score at all. Their key object is the Node-Edge tuple $s(i\to j)$5, aggregated as a multiset

$s(i\to j)$6

E-WL then refines node colors by hashing the current node color together with this node-edge tuple multiset (Yue et al., 4 Dec 2025). The result is edge-aware refinement, but not edge ranking. The closest analogue to an implicit edge contribution is the encoded tuple contribution inside the sum aggregator.

In HMGAT for hybrid beamforming, the closest edge-related scores are attention compatibilities. Node-level message passing uses

$s(i\to j)$7

while edge-level message passing uses

$s(i\to j)$8

with softmax-normalized coefficients $s(i\to j)$9 and $E_t$00 (Li et al., 10 Nov 2025). The same paper also uses “score-based” in the sense of $E_t$01 for CSI generation and denoising, showing that the word “score” there spans two distinct mathematical objects.

Two further cases delimit the concept. In generalized score matching for causal discovery, the score function is used to compute node-level leaf discriminants such as

$E_t$02

and the identified order is then used to constrain edge pruning or insertion; there is no direct per-edge score $E_t$03 (Vo et al., 22 Jan 2026). In Acc3D, the relevant quantity is the diffusion score function and its endpoint refinement through edge consistency in a high-SNR interval, not a score assigned to graph or image edges (Liu et al., 20 Mar 2025). These cases make clear that “Edge Score” is often best treated as a local term of art whose meaning must be recovered from the formal definition in the relevant paper, rather than inferred from the phrase alone.