Edge-Weighted QBAFs

• EW-QBAFs are quantitative bipolar argumentation frameworks that parameterize both arguments and relations using explicit edge weights for precise strength modulation.
• The framework employs modular and gradual semantics with sum and product aggregations to compute argument strengths, proving tractable and differentiable in acyclic graphs.
• By attributing relation-specific magnitudes, EW-QBAFs distinguish themselves from node-weighted models, enabling applications in forecasting, verification, and contestability analysis.

Searching arXiv for recent and foundational papers on EW-QBAFs and closely related QBAF semantics. Using arXiv search tool with focused queries on EW-QBAFs, QBAFs, modular semantics, and explainability. Edge-Weighted Quantitative Bipolar Argumentation Frameworks (EW-QBAFs) are quantitative bipolar argumentation models in which both arguments and relations are quantitatively parameterized. In the explicit formulation used for contestability, an EW-QBAF is a tuple

$$\mathcal{Q}=\left\langle\mathcal{A}, \mathcal{R}^{-}, \mathcal{R}^{+}, \tau, w \right\rangle$$

where $\mathcal{A}$ is a finite set of arguments, $$\mathcal{R}^{-}\subseteq\mathcal{A}\times\mathcal{A}$$ is the attack relation, $$\mathcal{R}^{+}\subseteq\mathcal{A}\times\mathcal{A}$$ is the support relation, $\tau:\mathcal{A}\rightarrow[0,1]$ assigns a base score to each argument, and $$w:\mathcal{R}^{-}\cup\mathcal{R}^{+}\rightarrow[0,1]$$ assigns an edge weight to each attack or support link (Yin et al., 15 Jul 2025). This distinguishes EW-QBAFs from the larger body of node-weighted QBAF and weighted bipolar graph research, in which attack and support edges are typically unweighted and only arguments carry quantitative information (Mossakowski et al., 2018).

1. Formalization and core objects

The defining feature of an EW-QBAF is that the contribution of a parent argument to a child argument is modulated by the weight of the connecting relation. For an argument $\alpha$, the incoming attacks and supports are written as

$$\mathcal{R}^{-}(\alpha) = \{ (\beta, \alpha) \mid (\beta, \alpha) \in \mathcal{R}^{-} \}, \qquad \mathcal{R}^{+}(\alpha) = \{ (\beta, \alpha) \mid (\beta, \alpha) \in \mathcal{R}^{+} \},$$

and the combined incoming relations as $$\mathcal{R}(\alpha)=\mathcal{R}^{-}(\alpha)\cup\mathcal{R}^{+}(\alpha)$$ (Yin et al., 15 Jul 2025). A gradual semantics then assigns each argument a final strength

$\sigma: \mathcal{A} \rightarrow [0,1]\cup\{\bot\},$

where $\mathcal{A}$0 is undefined only if $\mathcal{A}$1 (Yin et al., 15 Jul 2025).

The explicit edge-weighted formulation introduces weighted attacker and supporter multisets. Given a temporary strength function $\mathcal{A}$2, these are

$\mathcal{A}$3

$\mathcal{A}$4

This is the decisive mathematical difference from node-weighted QBAFs: relation intensity enters directly into propagation, rather than being inferred indirectly from the source node alone (Yin et al., 15 Jul 2025).

Much earlier work on bipolar quantitative argumentation uses the same basic attack/support architecture but without explicit edge magnitudes. In weighted attack/support argumentation graphs, for example, the graph matrix satisfies $\mathcal{A}$5, so an edge records only support, attack, or absence; the quantitative component is the initial-weight vector on arguments (Mossakowski et al., 2016). Likewise, modular semantics for bipolar weighted argumentation graphs uses an aggregation function and an influence function, but again with sign-only relations rather than weighted relations (Mossakowski et al., 2018). EW-QBAFs therefore retain the bipolar and gradual character of this literature while relocating part of the quantitative structure from nodes to edges.

2. Gradual semantics and computation

The main semantics developed explicitly for EW-QBAFs are modular. They factor strength computation into an aggregation stage over weighted attackers and supporters and an influence stage combining that aggregate with the base score $\mathcal{A}$6 (Yin et al., 15 Jul 2025). Two aggregation functions are given.

The first is sum aggregation: $\mathcal{A}$7

The second is product aggregation: $\mathcal{A}$8

The influence-function schema used for DF-QuAD and QE is

$\mathcal{A}$9

with

$$\mathcal{R}^{-}\subseteq\mathcal{A}\times\mathcal{A}$$0

for DF-QuAD and

$$\mathcal{R}^{-}\subseteq\mathcal{A}\times\mathcal{A}$$1

for QE (Yin et al., 15 Jul 2025). The resulting semantics are Edge-Weighted DF-QuAD, QE, REB, and MLP-based semantics, all presented as modular weighted variants (Yin et al., 15 Jul 2025).

The standard iterative computation initializes strengths at base scores,

$$\mathcal{R}^{-}\subseteq\mathcal{A}\times\mathcal{A}$$2

and updates them recursively using the weighted multisets $$\mathcal{R}^{-}\subseteq\mathcal{A}\times\mathcal{A}$$3 and $$\mathcal{R}^{-}\subseteq\mathcal{A}\times\mathcal{A}$$4 (Yin et al., 15 Jul 2025). For acyclic EW-QBAFs, this reduces to a topological forward pass, so strengths are computable in linear time; the same paper proves that, in this acyclic setting, the strength function is differentiable with respect to edge weights under Edge-Weighted QE, REB, DF-QuAD, and MLP-based semantics (Yin et al., 15 Jul 2025).

Adjacent QBAF work has refined the semantics side without explicitly adding edge weights. “Aggregative semantics” separates attacker aggregation, supporter aggregation, and final aggregation into three distinct stages,

$$\mathcal{R}^{-}\subseteq\mathcal{A}\times\mathcal{A}$$5

and explicitly observes that weighted binary relations can be added by “keeping the same formalism” while modifying the domains of the aggregation functions (Munro et al., 6 Mar 2026). Likewise, Double Rectified Linear Unit-based modular semantics introduces a normalized influence term $$\mathcal{R}^{-}\subseteq\mathcal{A}\times\mathcal{A}$$6 and a bounded update

$$\mathcal{R}^{-}\subseteq\mathcal{A}\times\mathcal{A}$$7

but still within an unweighted-edge QBAF syntax (Alfano et al., 4 May 2026). This suggests that, in current theory, the aggregation layer is the natural insertion point for edge-weight information.

3. Relation to node-weighted bipolar argumentation

A recurring theme in the literature is that many frameworks closely related to EW-QBAFs are not actually edge-weighted. Bipolar weighted argumentation graphs, weighted attack/support argumentation graphs, and BAGs all place weights on arguments while representing relations only by polarity (Mossakowski et al., 2016, Potyka, 2018). In these models, support and attack are often aggregated through a signed sum, a top operator, or a product operator, but no relation-specific magnitude is present (Potyka, 2018).

This distinction matters both semantically and computationally. In node-weighted models, a supporter’s impact depends on its current strength and its support status; in EW-QBAFs, the same supporter can contribute differently to different targets because the weight is attached to the edge, not solely to the source node. The explicit weighted multisets $$\mathcal{R}^{-}\subseteq\mathcal{A}\times\mathcal{A}$$8 and $$\mathcal{R}^{-}\subseteq\mathcal{A}\times\mathcal{A}$$9 in EW-QBAFs formalize exactly this target-sensitive contribution (Yin et al., 15 Jul 2025).

A common misconception is that practical QBAF systems used in forecasting or retrieval-augmented verification already implement edge-weighting. They typically do not. In the multi-agent forecasting framework based on QBAFs, quantitative information lives primarily on arguments via base scores $$\mathcal{R}^{+}\subseteq\mathcal{A}\times\mathcal{A}$$0, while support and attack edges are only typed; the paper explicitly notes that there is no edge-weight function $$\mathcal{R}^{+}\subseteq\mathcal{A}\times\mathcal{A}$$1 and no confidence score attached to relations themselves (Gorur et al., 28 Oct 2025). Similarly, ArgRAG defines a QBAF as

$$\mathcal{R}^{+}\subseteq\mathcal{A}\times\mathcal{A}$$2

with base scores on arguments and binary support/attack relations, and its QE aggregation uses unweighted sums over supporters and attackers (Zhu et al., 26 Aug 2025). From an EW-QBAF standpoint, these systems are best viewed as unweighted-edge baselines or application templates rather than full edge-weighted formalisms.

4. Attribution, sensitivity, and contestability

EW-QBAFs have been studied directly in the context of contestability. The contestability problem asks how to modify edge weights so that a chosen topic argument attains a desired final strength (Yin et al., 15 Jul 2025). Formally, given a topic argument $$\mathcal{R}^{+}\subseteq\mathcal{A}\times\mathcal{A}$$3 and target $$\mathcal{R}^{+}\subseteq\mathcal{A}\times\mathcal{A}$$4, the task is to find a new edge-weight function $$\mathcal{R}^{+}\subseteq\mathcal{A}\times\mathcal{A}$$5 such that

$$\mathcal{R}^{+}\subseteq\mathcal{A}\times\mathcal{A}$$6

for

$$\mathcal{R}^{+}\subseteq\mathcal{A}\times\mathcal{A}$$7

Under edge-stability, the base score $$\mathcal{R}^{+}\subseteq\mathcal{A}\times\mathcal{A}$$8 is always attainable, and under edge-monotonicity, monotonicity, and edge-neutrality, the attainable strengths form an interval $$\mathcal{R}^{+}\subseteq\mathcal{A}\times\mathcal{A}$$9 provided the semantics is continuous (Yin et al., 15 Jul 2025).

The central explanatory device is the Gradient-based Relation Attribution Explanation (G-RAE),

$\tau:\mathcal{A}\rightarrow[0,1]$0

where only the weight of relation $\tau:\mathcal{A}\rightarrow[0,1]$1 is perturbed (Yin et al., 15 Jul 2025). The sign of $\tau:\mathcal{A}\rightarrow[0,1]$2 indicates whether increasing that edge weight raises, lowers, or leaves unchanged the topic strength. For direct edges, support relations have nonnegative influence and attack relations have nonpositive influence under edge-monotonicity; for indirect edges, the sign flips with the parity of attacks along the path (Yin et al., 15 Jul 2025). If a relation is independent of the topic, the derivative is $\tau:\mathcal{A}\rightarrow[0,1]$3 under edge-directionality (Yin et al., 15 Jul 2025).

The same paper proves tractability: exact G-RAEs can be generated in linear time $\tau:\mathcal{A}\rightarrow[0,1]$4 for acyclic EW-QBAFs, and a perturbation-based approximation computes all approximate G-RAEs in time

$\tau:\mathcal{A}\rightarrow[0,1]$5

Building on these attributions, the contestability algorithm iteratively updates all edge weights by projected gradient steps

$\tau:\mathcal{A}\rightarrow[0,1]$6

until the target is reached within tolerance or the iteration budget is exhausted; its time complexity is

$\tau:\mathcal{A}\rightarrow[0,1]$7

for acyclic EW-QBAFs (Yin et al., 15 Jul 2025). Experiments on synthetic personalised recommender system-like and multilayer perceptron-like EW-QBAFs report $\tau:\mathcal{A}\rightarrow[0,1]$8 validity across the tested settings, with average runtime under $\tau:\mathcal{A}\rightarrow[0,1]$9 seconds on the former and maximum average runtime $$w:\mathcal{R}^{-}\cup\mathcal{R}^{+}\rightarrow[0,1]$$0 seconds on the latter (Yin et al., 15 Jul 2025).

A related but distinct line of work studies relation-level explanation in non-edge-weighted QBAFs. Relation Attribution Explanations (RAEs) assign post hoc importance scores to edges,

$$w:\mathcal{R}^{-}\cup\mathcal{R}^{+}\rightarrow[0,1]$$1

or their Shapley-based analogue, but these scores are explanatory quantities rather than semantic edge weights (Yin et al., 2024). This distinction is essential: an important edge in an explanation is not automatically a weighted edge in the underlying semantics.

5. Application contexts and empirical templates

The most developed applications around QBAFs have so far used node-weighted rather than edge-weighted graphs, but they indicate where EW-QBAFs would be especially useful. In judgmental forecasting, different agents produce QBAFs for the same claim and a multi-agent combinator merges semantically similar arguments while aggregating node base scores; the framework does not use edge weighting, and the paper explicitly identifies as missing from an EW-QBAF perspective an explicit edge-weight function, semantics using edge magnitudes in propagation, retrieval-confidence weighting, agent-reliability weighting, and conflict-resolution mechanisms for relation confidence (Gorur et al., 28 Oct 2025).

ArgRAG provides a second clear template. It constructs a QBAF from retrieved evidence passages and claim/evidence or evidence/evidence support and attack relations, then applies QE gradual semantics for deterministic inference (Zhu et al., 26 Aug 2025). The retrieved passages become arguments, and all arguments are initialized with uniform base score $$w:\mathcal{R}^{-}\cup\mathcal{R}^{+}\rightarrow[0,1]$$2; retrieval scores and relation confidences are not propagated as edge weights (Zhu et al., 26 Aug 2025). The paper explicitly notes that a natural EW-QBAF reformulation would replace the unweighted aggregate

$$w:\mathcal{R}^{-}\cup\mathcal{R}^{+}\rightarrow[0,1]$$3

by a weighted version with relation-specific coefficients (Zhu et al., 26 Aug 2025). This suggests that high-stakes retrieval settings are a natural application domain for EW-QBAFs, precisely because evidential relevance, contradiction confidence, and source credibility are inherently relation-specific.

Forecasting Argumentation Frameworks provide another adjacent case: they use typed proposal, amendment, pro, and con arguments, with DF-QuAD-derived strengths over an acyclic graph, but no edge weights (Irwin et al., 2022). Here too, quantitative argumentation constrains downstream forecasts, while the revision intensity associated with different argumentative links remains unweighted.

6. Structured, dynamic, and open extensions

Several recent directions extend quantitative bipolar argumentation beyond flat node-weighted graphs and are directly relevant to future EW-QBAF theory. A methodology for incompleteness-tolerant gradual semantics on statement graphs lifts QBAF semantics such as DF-QuAD and QEM into a structured setting where support and attack are induced by premise-claim matching; the framework still uses node weights only, but it explicitly presents itself as a bridge from abstract QBAF semantics to premise-sensitive structured argumentation (Rago et al., 2024). In assumption-based argumentation, modular QBAF semantics has been generalized to bipolar set-based frameworks with collective attacks and supports, introducing a three-layer architecture—set aggregation $$w:\mathcal{R}^{-}\cup\mathcal{R}^{+}\rightarrow[0,1]$$4, relation aggregation $$w:\mathcal{R}^{-}\cup\mathcal{R}^{+}\rightarrow[0,1]$$5, and influence $$w:\mathcal{R}^{-}\cup\mathcal{R}^{+}\rightarrow[0,1]$$6—that is structurally very close to a hyperedge-weighted generalization of EW-QBAFs (Rapberger et al., 14 Jul 2025).

Dynamic change is another active direction. Quantitative argumentation dialogues have been studied as sequences of quantitative bipolar graphs equipped with temporal notions of safety, liveness, and fairness over argument strengths, again with node weights rather than edge weights (Ganguly et al., 22 May 2026). Explanations of change across successive QBAFs have also been developed in terms of sufficient, necessary, and counterfactual explanations for changes in the partial order over topic-argument strengths, with reversals defined over arguments, base scores, and outgoing attack/support relations (Kampik et al., 21 Sep 2025). This suggests that a full EW-QBAF theory of change would need reversal and explanation operators that restore not only relation structure but also relation weights.

Across these extensions, one consistent pattern emerges. Theoretical infrastructure for gradual semantics, aggregation, structured lifting, temporal monitoring, and explanation is already substantial, but much of it remains formulated for node-weighted bipolar graphs. Explicit edge weighting has, by the wording of the contestability paper, “received little attention” (Yin et al., 15 Jul 2025). The main open directions repeatedly indicated by adjacent work are the integration of relation confidence into propagation, the extension of semantic postulates and convergence theory to weighted relations, and the migration from argument-level explanation to genuinely edge-sensitive explanation and intervention (Gorur et al., 28 Oct 2025, Munro et al., 6 Mar 2026).