Bipolar Weighted Argumentation Graphs

• Bipolar weighted argumentation graphs are formal frameworks that combine positive (support) and negative (attack) relations with numerical weights to quantify argument strength.
• They employ modular, aggregative, and continuous semantic approaches, using iterative fixed-point equations or dynamical systems to compute graded acceptability.
• These frameworks support contestable, explainable reasoning in AI applications such as decision support, forecasting, and personalized recommendation systems.

Bipolar weighted argumentation graphs generalize classical argumentation theory by integrating both positive (support) and negative (attack) relations between arguments, combined with numeric quantification on nodes and/or edges. Their formalism supports graded, explainable aggregation of preference, plausibility, or evidential strength across complex networks, and underpins a diverse landscape of semantics, axiomatizations, and algorithmic frameworks for reasoning and explainability in AI, preference aggregation, and decision support.

1. Formal Foundations and Definitions

A bipolar weighted argumentation graph (BWAG) is formally defined as a tuple encoding a finite set of arguments, both attack and support relations, and a numeric structure that quantifies node and/or edge strength. The canonical definition, following (Mossakowski et al., 2016), is: $\mathcal{A} = (A, G, w)$ where $A$ is a finite set of arguments, $G \in \{-1, 0, +1\}^{n \times n}$ the incidence matrix ($g_{ij}=+1$ if $a_j$ supports $a_i$, $g_{ij}=-1$ if $a_j$ attacks $a_i$), and $w \in \mathbb{R}^n$ assigns an initial plausibility score to each argument. Extensions and variants, notably the Edge-Weighted Quantitative Bipolar Argumentation Framework (EW-QBAF) (Yin et al., 15 Jul 2025), further assign a weight $w:\mathcal{R}^-\cup\mathcal{R}^+\to[0,1]$ to each support/attack edge.

Attack and support relations are defined as:

• $\mathcal{R}^- \subseteq A \times A$ (attack edges)
• $\mathcal{R}^+ \subseteq A \times A$ (support edges)
• $\mathcal{R}^- \cap \mathcal{R}^+ = \emptyset$

Node weights ($\tau: A \to [0,1]$) provide initial acceptability, plausibility, or preference bases.

2. Semantics: Modular, Aggregative, and Continuous Approaches

Quantitative evaluation in BWAGs is governed by gradual semantics, which assign real-valued acceptability degrees to arguments via iterative or fixed-point equations. The modular semantics framework (Mossakowski et al., 2018) decomposes evaluation into two core functions:

1. Aggregation function ($\alpha$): Merges strengths of attacking and supporting parents.
2. Influence function ($\iota$): Modulates the initial weight in response to the aggregated parent strength.

The general update for argument $a_i$ is: $D_i = \iota(\alpha(G_{i,*}, D), w_i)$ Sum-based ($\alpha_{\mathrm{sum}}$), product-based, and max-based ($\alpha_{\mathrm{top}}$) aggregators are typical choices, with corresponding influence functions (linear, Euler-based, or non-linear).

The direct aggregation semantics, introduced in (Mossakowski et al., 2016), employs a fixed-point equation: $s^d = w + \frac{1}{d} G s^d$ or equivalently,

$s^d = (I - \frac{1}{d}G)^{-1} w$

requiring $d > \mathrm{max\ indegree}(G)$ for convergence.

Aggregative semantics (Munro et al., 6 Mar 2026) further disaggregate attacker and supporter contributions, computing them via separate functions and combining with the intrinsic strength through a final aggregation, supporting asymmetric influence of positive and negative edges.

Continuous dynamical system models (Potyka, 2018) represent argument strengths $s(t)$ as time-evolving ODEs, e.g., for the quadratic energy model: $$\dot{s}_j = w(j) - s_j + (1-w(j))h(E_j) - w(j)h(-E_j)$$ with $h(x) = \max\{x,0\}^2 / (1+\max\{x,0\}^2)$, offering improved stability in cyclic graphs.

3. Gradual Evaluation Workflow and Convergence

Evaluation typically proceeds through:

1. Initialization: Set strengths to node base weights.
2. Iterative updates: At each iteration (discrete semantics) or continuously (continuous semantics), aggregate influences, apply the influence function, and update argument strengths.
3. Fixed-point/convergence: Under acyclicity or contraction conditions, the process converges to stable argument strengths.

Convergence is guaranteed for acyclic graphs under almost all semantics (Mossakowski et al., 2018). For cyclic graphs, convergence may fail for sum-based semantics but is ensured for max-based variants ($2m < 1$ for the Lipschitz constant of influence functions), or via the continuous ODE approach (Mossakowski et al., 2018, Potyka, 2018). Runtime for acyclic graphs is linear or polynomial in graph size, depending on the evaluation scheme.

4. Edge Weight Contestability and Explainability

A distinguishing feature of edge-weighted frameworks, notably EW-QBAF (Yin et al., 15 Jul 2025), is their support for contestability: modifying edge weights to achieve desired outcome strengths for a focal argument. Formally, given $\sigma(\alpha)$ and target $s \neq \sigma(\alpha)$, the task is to find a new edge weight function $w'$ with $\sigma_{w'}(\alpha) = s$.

Gradient-based Relation Attribution Explanations (G-RAEs) quantify the influence of an edge's weight on the strength of a target argument, computed as

$$\nabla_{r \to \alpha}^\sigma = \frac{\partial \sigma(\alpha)}{\partial w(r)}$$

Positive gradients indicate edges whose increased weight would raise $\sigma(\alpha)$; negative gradients mark detrimental edges. This supports interpretable, gradient-based edge weight adjustment, operationalized as a projected gradient descent scheme to achieve desired strengths.

These approaches provide local sensitivity and explainability analogous to backpropagation in neural networks, demonstrating computational efficiency and transparent, controllable AI behavior (Yin et al., 15 Jul 2025).

5. Axiomatic Properties and Semantics Comparison

A robust axiomatization underpins soundness and rationality. Key postulates include:

• Anonymity: Argument labels do not affect outcomes.
• Directionality: Only reaching arguments matter.
• Stability: No parents yields base strength.
• Monotonicity, reinforcement, neutrality, continuity
• Counterfactuality: Removal effect interpretation.
• Additivity (Shapley-based): Contributions sum to the global change in acceptability.

Table: Satisfiability of key principles by contribution function classes (Kampik et al., 2024, Naudot et al., 18 Sep 2025):

Contribution Function	Counterfactuality	Additive	Local Sensitivity (Faithfulness)
Removal-based	✔	✗	✗
Intrinsic-removal	✗	✗	✗
Shapley-based	✗	✔	✗
Gradient-based	✗	✗	✔

No single function satisfies all user-desired axiomatic constraints; selection depends on application priorities (counterfactuality, additivity, sensitivity) (Kampik et al., 2024, Naudot et al., 18 Sep 2025).

6. Applications and Extensions

BWAGs provide an expressive substrate for numerous AI domains:

• Personalized decision support and contestable AI: Edge-weight tuning aligns system output with human-centric goals (Yin et al., 15 Jul 2025).
• Multi-user preference aggregation: QBAFs and their multi-user extensions resolve conflicting stakeholder preferences, as in human-robot interaction decision mediation (Civit et al., 5 Nov 2025).
• Forecasting frameworks: Used in group forecasting, combining weighted bipolar arguments with agent-based constraints and rationality principles (Irwin et al., 2022).
• Explainable recommendation systems: Contribution functions (removal-based, gradient-based, Shapley) enable fine-grained, multi-aspect explanation in recommender scenarios (Naudot et al., 18 Sep 2025).

Connections to neural network analysis are formalized via identification of arguments with network nodes and edge weights with synaptic strengths, with argument equilibria corresponding to network stationary states (Giordano, 2021).

7. Perspectives and Future Directions

Recent research has dramatically expanded the formal and practical reach of BWAGs. Open challenges include:

• General convergence in the presence of cycles: While sum- and product-based modular semantics may diverge, max-based and continuous dynamical approaches exhibit strong convergence guarantees (Mossakowski et al., 2018, Potyka, 2018).
• Edge versus node weighting: The landscape includes node-only, edge-only, and joint weighting schemes; their empirical and axiomatic ramifications merit systematic comparison (Mossakowski et al., 2016, Yin et al., 15 Jul 2025, Giordano, 2021).
• Expanding axiomatics and contribution analysis: Extension of single-argument to set-focused contribution analysis, together with new interaction principles and principle-based selection frameworks, advances the transparency and robustness of explanations (Naudot et al., 18 Sep 2025).
• Implementation frameworks: Libraries such as Attractor support rapid prototyping and benchmark-oriented exploration of semantic and algorithmic design space (Potyka, 2018, Potyka, 2018).

Through modular design, robust axiomatics, and algorithmic advances, bipolar weighted argumentation graphs have become a foundational paradigm for quantitative, explainable, and contestable reasoning in AI.