Argument Reweighting in WAFs

• Argument reweighting is a method that systematically refines argument weights in WAFs using interval-based elicitation and rationality constraints.
• It employs gradual semantics—such as h-Categorizer, Card-Based, and Max-Based—to compute final acceptability degrees via recursive equations.
• The approach enhances consistency by iteratively refining user-supplied intervals and minimally repairing irrational configurations for robust evaluations.

Argument reweighting is a formal methodology for eliciting, refining, and correcting the initial quantitative strengths (“weights”) assigned to arguments within weighted argumentation frameworks (WAFs). This process is necessitated by the empirical difficulty of specifying precise initial weights and the confusion that often arises between these weights and the arguments’ derived final acceptability degrees. Argument reweighting leverages interval-based elicitation, rationality constraints, and gradual semantics to systematically map plausible acceptability intervals into consistent initial weight vectors, enabling robust, principled interaction with argumentative structures (Oren et al., 11 Feb 2025).

1. Weighted Argumentation Frameworks (WAFs)

A weighted argumentation framework is formally defined as a triple $(A, C, w)$, where:

• $A = \{a_1, \dots, a_n\}$ is a finite set of arguments,
• $C \subseteq A \times A$ is the attack relation,
• $w: A \to [0, 1]$ assigns an initial weight to each argument.

Two critical notational constructs enhance computational manipulation:

• $Att(a) = \{b \in A \mid (b, a) \in C\}$, representing direct attackers of $a$,
• $Att^*(a) = \{b \in Att(a) \mid w(b) \neq 0\}$, restricting to nonzero-weight attackers.

The weights $w(a)$ serve as the foundational input to ranking-style semantics that ultimately produce acceptability degrees for each argument.

2. Gradual Semantics and Their Role

Gradual semantics define recursive transformation from initial weights to final acceptability degrees $\sigma(a) \in [0,1]$. The prominent semantics covered are:

• Weighted h-Categorizer:

$$\sigma(a) = \frac{w(a)}{1 + \sum_{b \in Att(a)} \sigma(b)}$$

• Weighted Card-Based:

$$\sigma(a) = \begin{cases} \frac{w(a)}{1 + |Att^*(a)| + \frac{1}{|Att^*(a)|} \sum_{b \in Att^*(a)} \sigma(b)} & Att^*(a) \neq \emptyset \ w(a) & Att^*(a) = \emptyset \end{cases}$$

• Weighted Max-Based:

$$\sigma(a) = \frac{w(a)}{1 + \max_{b \in Att(a)} \sigma(b)}$$

For each framework and semantics, one solves a system of nonlinear equations to obtain a unique degree vector. These degrees—representing argument acceptability—crucially depend on both the graph topology and initial weights.

3. Interval-Based Elicitation Pipeline

Precise weight input is often impractical; the proposed protocol solicits intervals $I(a) = [L_a, U_a] \subseteq [0, 1]$ for each argument, capturing plausible acceptability degrees. The pipeline seeks to guarantee rationality, i.e., the existence of some initial weighting $w$ such that $\sigma_w(a) \in I(a)$ for all $a$, and to refine intervals if necessary.

Key concepts include:

• Rationality: $(cAF, \sigma)$ is rational if some $w$ yields $\sigma_w(a)\in I(a)$ for all $a$.
• Full Rationality: Achieved if every degree vector $d(a)\in I(a)$ is realizable.

For h-Categorizer and Max-Based semantics, rationality checking focuses on whether the “minimal corner” $(L_1,\dots,L_n)$ sits within the acceptability-degree space. The interval refinement (Algorithm 1) iteratively tightens each $U_a$ using binary search and rationality-oracle calls until no substantial further tightening is possible. This process executes in $O(n\log(1/\varepsilon))$ rounds, leveraging degree-space monotonicity.

4. Restoring Rationality in Irrational Configurations

When user-supplied intervals are irrational (i.e., no $w$ yields all $\sigma_w(a)\in I(a)$), a minimal-cost repair is constructed by “nudging” the lower bounds $L_a$ upward. This is achieved by:

• Assigning cost factors $c(a)>0$,
• Defining a CAF family $I_t(a) = [L_a + t/c(a), U_a]$ for $t\ge 0$,
• Binary-searching the minimal $t$ needed for rationality, implementing a weighted $L_\infty$ correction.

An alternative “Strategy 2” allows for subset selection and localized repairs but is computationally infeasible for large $n$ due to combinatorial explosion. This methodology ensures interval consistency at the least aggregate perturbation, thereby preserving fidelity to user intent.

5. Derivation and Recovery of Initial Weights

Upon achieving a rational configuration $(A, C, I)$, one recovers the concrete initial weights $w$ corresponding to the interval-elicitated degrees:

• Select a target degree vector $d(a)\in I(a)$ (commonly mid-point sampling),
• Solve the inverse problem $\sigma_w(a) = d(a)$.

For h-Categorizer, analytical inversion yields:

$w(a) = d(a) + d(a) \sum_{b \in Att(a)} d(b)$

For other semantics, matrix-based reconstruction methods (cf. Oren et al., 2022) are applicable. Sampling within intervals allows exploration of heterogeneity in plausible weight assignments, supporting robustness analyses.

6. Illustrative Worked Example

Consider $A = \{a, b\}$, $C = \{(a, b)\}$, semantics = h-Categorizer, and user intervals $I(a) = [0.8, 1]$, $I(b) = [0.5, 0.6]$. The acceptability space is bounded by $d_a = 1/(1 + d_b)$:

• If $I(b) = [0.5, 5/9] \approx [0.5, 0.5556]$, the CAF is fully rational (every corner lies under the bounding curve).
• If $I(b) = [0.5, 0.7]$, rational but not fully rational; refinement is required to $U_b = 5/9$.
• If $I(b) = [0.6, 0.8]$, irrational; the cost-nudge method adjusts $L_a$ to the minimal feasible value.

This demonstrates the pipeline's capacity to interactively enforce rationality and extract compatible initial weights.

7. Comparative Context and Methodological Significance

Relative to existing approaches:

Approach	Core Capability	Interval Support
Inverse Gradual Semantics (Oren+)	Solves $\sigma_w(a) = d(a)$	No (exact degrees)
Credal/Bayesian Support AFs	Distributions over beliefs	Not tied to WAF
Constraint-Based Argumentation	CSPs for final degrees	No (exact values)
Argumentation Enforcement (Baumann+)	Graph modification for acceptability	No (modifies graph)

Argument reweighting introduced here allows interval-based believability, interactive tightening, and minimal-change rationalization—distinct from the exact-degree or graph-editing protocols. Notable advantages include eased elicitation (intervals, not numbers), automatic space-tightening, cost-aware inconsistency repair, and direct recovery of initial weights consonant with user input and argumentative structure.

Limiting factors include computational demands of rationality oracle calls and combinatorial limits of subset selection strategies. The extension to distributions over degrees remains open.

In summary, argument reweighting provides a formal, interactive, and principled pipeline for obtaining rational, minimally-adjusted initial weights in weighted argumentation frameworks given interval-based user inputs, underpinned by gradual semantics and robust rationality constraints (Oren et al., 11 Feb 2025).