acceptodds
Under review as a conference paper at ICLR 2027

Vector-Valued Quantitative Bipolar Argumentation for Node Classification

Abstract

Node classification combines evidence from node attributes with potentially conflicting information from graph neighbors. We develop an end-to-end trainable architecture based on vector-valued quantitative bipolar argumentation frameworks (QBAFs). Feature-grounded arguments carry class-specific evidence, while dynamic support, neutral, and attack relations determine how that evidence revises neighboring arguments and class claims. Individual and jointly composed evidence contribute through the same role algebra, and prediction uses only the accumulated terminal claims. These operations expose feature contributions, relational actions, and score increments within the predictive computation itself. We prove that single-channel role interventions have class-specific spatial effects and derive downstream score bounds with conditions for preserving the predicted class. Against 16 comparison methods, QBAF achieves the highest tabulated mean on a majority of the benchmarks—six of nine datasets. Forty-eight matched ablation runs establish the contribution of local argument composition and measure the effects of relational conditioning and class weighting. Recorded traces reconstruct predictions, and controlled interventions identify evidence sustaining correct decisions or contributing to errors. The architecture connects graph learning and quantitative argumentation through a prediction mechanism whose reasoning can be reconstructed and intervened upon.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
graph representation learning, node classification, quantitative bipolar argumentation, interpretable deep learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_LFi4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Inspired by the concepts in Graph Learning and Computational Argumentation, this paper proposes an explainable graph learning model, QBAF, which treats the Node Classification task as an Agreement Classification task. The interpretability comes from transforming the implicit node communication in traditional methods into an explicit process of debate. It starts with node features generating arguments with evidence on each class. These arguments interact within nodes and across edges through learned support, neutral, and attack roles, collectively contributing to the update of the class scores. The paper evaluates predictive performance on small scales Node Classification datasets, shows how recorded computations reconstruct predictions, and provides theoretical analysis of the proposed framework.

Critical strengths

1.The idea bridging the two fields is good though I am not familiar with Computational Argumentation. 2.Theoretical Analysis of the proposed framework is thorough. The scores reconstruction, role intervention examination, and bound analysis seems solid. 3.The empirical results look promising. Best in 6 out of 9 datasets competing with 16 recent baselines.

Critical weaknesses

1.It seems that many baseline results were reused from existing/original papers, for example the Specformer. Reusing published baseline results is acceptable only if their sources are clear and comparability is fair. So it is justifiable to ask: Which results were rerun and which were reused? What are the sources of the resued numbers? Are data splits, preprocessing, and tuning procedures comparable? What each reported uncertainty measure represents? 2.QBAF can be quite expensive to scale. Although Appendix A.6 provides complexity analysis and Section 5.4 reports some timings, the overhead on larger graphs remains unclear. For example, one of the largest citation network ogbn-papers100M has 111,059,956 nodes and 172 classes. Assuming 64 arguments per class and FP32 precision, it results in 4.89TB memory for one full state tensor. It is inappropriate for me to demand such large-scale datasets, perhaps a small-scale dataset like ogbn-arxiv is a more reasonable request. How expensive is QBAF in practice? Could you measure peak memory and training/inference time alongside accuracy? 3.The “Support, Neutral, and Attack” framing sounds intuitive. My concern is: an interaction with (0.45S, 0.10N, 0.45A) gives quite different role assignments compared to (0.05S, 0.9N, 0.05A), yet both contribute zero according to Eq.5. Indeed, the roles are mathematically well-defined, but it seems that understanding their effect still requires us to look at the signed message and the later updates. Could you extend the existing cases to show what the role probabilities reveal beyond the net signed coefficient and payload? 4.It seems that the classes share features and some parameters and are trained together but they do not directly use each other’s evolving arguments/scores during reasoning. I understand that this can make interventions much easier but wouldn’t it limit how the model uses relationships between competing classes during reasoning? If this is a design trade-off then its impact needs to be tested.

Questions for authors

An interesting idea still needs developed motivation. How QBAF lets us understand or do beyond existing signed propagation and intrinsic explanation methods?

1 comment
AuthorsOpenReview ↗

Source-Tracked Baselines, Measured Scalability, Role-Sensitive Edits, and Cross-Class Reasoning

1. Reproducible baselines and matched comparisons (Critical Weakness 1). Tables 9–10 trace all 144 published-baseline cells and their uncertainty conventions, including Specformer from GrokFormer (ten-run SD) and PolyFormer (Questions: 95% interval). Table 1 separately identifies our ten QBAF fits. We additionally reran QBAF and Specformer on all nine graphs with matched features, node order, adjacency and splits, retaining model-specific normalization. Validation-only tuning searched 24 configurations on Photo/Actor and six elsewhere. Seven paired intervals favor QBAF in Table 11, including Photo +1.361 accuracy points [1.125, 1.598] and Questions +3.720 ROC-AUC points [2.675, 4.765]. Five designated-comparator gains survive ten-split sign-flip tests and Holm correction (Table 12).

2. White-box scalability and measured cost (Critical Weakness 2). We extended QBAF to ogbn-arxiv's official temporal split (169,343 nodes; 40 classes). Ten-layer QBAF achieves 71.53±0.3171.53\pm0.31\\% test accuracy across five initializations, exceeding GPR-GNN's 68.80% by 2.73 accuracy points (Tables 4, 32). White-box computation tracks 64 class-specific argument slots, support/neutral/attack roles and layerwise score revisions, enabling trace replay and targeted edits. This explicit computation adds per-pass cost: on RTX 5090 D, a full-graph update takes 2.150 s, inference 0.608 s, and peak allocated training memory is 23.840 GiB (Table 33). Class blocking, sparse propagation and activation recomputation preserve reference predictions at all five checkpoints, establishing the large-graph accuracy and cost of traceable argumentation.

3. Information beyond the net signed message (Critical Weakness 3). The reviewer's two triples (.45,.10,.45)(.45,.10,.45) and (.05,.90,.05)(.05,.90,.05) both give ν=2(pS−pA)=0\nu=2(p_S-p_A)=0. Equal and opposite support/attack-logit changes give sensitivity κ=2[pS+pA−(pS−pA)2]\kappa=2[p_S+p_A-(p_S-p_A)^2], equal to 1.8 and 0.2 for the two triples. On Photo edge 1160→131160\to13 (layer 2, head 0, class 1), both replacements yield the same starting raw margin 7.487156; a +0.1 tilt changes the fully recomputed margin by +0.022266 versus +0.002510 (Figure 1; Table 28). Four controlled cases confirm larger high-sensitivity responses in both directions (Table 27). Channel selection also uses downstream influence Je=∂M/∂νeJ_e=\partial M/\partial\nu_e. On 100 Photo targets, ranking up to three channels by ∣Jeκe∣|J_e\kappa_e| yields larger recomputed margin changes than ranking by ∣Je∣|J_e| or ∣Jeνe∣|J_e\nu_e| for both edit directions at both checkpoints (Table 26). Thus, roles predict intervention responses beyond the signed message.

4. Why class-separable updates matter (Critical Weakness 4). We compared the original class-separable model with a same-class capacity control, per-layer cross-class mixing and final-layer mixing (Tables 29–31). Per-layer mixing reads competing arguments and claims. On Actor, the original model attains 28.75 macro-F1 and exceeds per-layer mixing by 1.372 points [0.278, 2.465]; the extension increases update time by 16.11% and inference time by 46.30%. On WikiCS, per-layer mixing improves macro-F1 over final-layer mixing by 0.253 points [0.058, 0.447]. The class-separable program preserves other classes' raw scores under a single-class edit, while the shared softmax still connects their output probabilities. A controlled edit produces a maximum competing raw-score shift of 1.49×10−81.49\times10^{-8} in the base versus 1.69×10−41.69\times10^{-4} under per-layer mixing. This quantifies the predictive and intervention effects of class coupling.

5. What QBAF enables beyond signed propagation (Question for Authors). QBAF exposes feature-grounded arguments, role assignments and class-score revisions for exact replay and targeted edits, distinguishing equal-message relations by their intervention effects. In a matched direct-signed control holding arguments, local composition, claim feedback, readout and weighting fixed, three roles improve Cora accuracy by +1.478 points [0.879, 2.077] under EM and Actor macro-F1 by +1.959 points [0.726, 3.192] under unit weights (Table 16). The explanation study covers 600 targets, with 575 permitting non-self edits. On 100 Actor targets and the same fixed QBAF predictor, native evidence selection improves deletion and masking gap AUC over GNNExplainer by +0.005152 and +0.005647, and retention gap by -0.006265 [-0.009749, -0.003277] (Tables 19–20). Across complete Actor predictors, all 12 pointwise deletion/masking intervals favor QBAF over direct-signed, FAGCN and SES at two checkpoints (Table 21). These results establish white-box prediction, executable edits and measured explanation utility in one model.

Reviewer_mQE1OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a vector-valued quantitative bipolar argumentation framework for node classification. It represents node features as class-specific arguments and models interactions through learned support, neutral, and attack relations. Class predictions are obtained through accumulated argument revisions, allowing the computation to be traced and intervened upon. Experiments on nine graph datasets report competitive classification performance, supported by ablations and intervention analyses.

Critical strengths
  1. The paper explores an interesting connection between quantitative argumentation and graph learning. Incorporating explicit relational roles into prediction is a reasonable approach to intrinsic interpretability.

  2. The experiments cover nine datasets and several GNN and Graph Transformer baselines. The paper also provides component ablations and theoretical analysis of intervention effects.

Critical weaknesses
  1. The methodological distinction from signed message passing and attention-based GNNs is not sufficiently clear.

  2. The interpretability claims are insufficiently validated. Exact reconstruction of prediction scores follows from the model's additive formulation, but does not establish that the learned arguments and relational roles provide meaningful explanations.

  3. The intervention experiments are limited to four selected cases, with only three supporting non-self-edge interventions. The small control pools make it difficult to draw general conclusions about explanation quality.

  4. The ablation results do not consistently support the importance of the proposed relational mechanisms. Removing target conditioning produces little or no degradation on several datasets, while class reweighting accounts for substantial improvements on Cora and Texas.

  5. The predictive improvements are uneven. Several gains over the strongest baselines are small relative to the reported variability, and QBAF underperforms on Actor, Texas, and Questions.

Questions for authors
  1. Clarify the methodological difference from signed GNNs.

  2. Provide quantitative evaluation of explanation quality beyond score reconstruction.

  3. Explain the limited contribution of relational conditioning in the ablations.

  4. Provide stronger statistical support for the reported performance improvements.

3 comments
AuthorsOpenReview ↗

Role-Dependent Edits, Explanation Fidelity, Matched Ablations, and Predictive Gains

1. Role-specific graph computation (Critical Weakness 1; Question 1). QBAF propagates class-specific argument evidence with me=αeνeuem_e=\alpha_e\nu_e u_e and νe=2(pS,e−pA,e)\nu_e=2(p_{S,e}-p_{A,e}). The role distribution adds a measurable response to a specified operation. Under a support/attack logit tilt (+ϵ,−ϵ)(+\epsilon,-\epsilon), κe=∂νe∂ϵ∣ϵ=0=2[pS,e+pA,e−(pS,e−pA,e)2]\kappa_e=\left.\frac{\partial\nu_e}{\partial\epsilon}\right|_{\epsilon=0}=2[p_{S,e}+p_{A,e}-(p_{S,e}-p_{A,e})^2], and the margin response is JeκeJ_e\kappa_e, where Je=∂M/∂νeJ_e=\partial M/\partial\nu_e. On Photo, (.45,.10,.45)(.45,.10,.45) and (.05,.90,.05)(.05,.90,.05) give identical forward messages and the same starting margin; a +0.1+0.1 edit changes that margin by 0.0222660.022266 versus 0.0025100.002510 under full recomputation (Tables 27–28). On 100 Photo targets, ∣Jeκe∣|J_e\kappa_e| selects edits with larger measured effects than ∣Je∣|J_e| or ∣Jeνe∣|J_e\nu_e| in both directions at two checkpoints (Table 26). GATv2 normalizes attention among neighbors; QBAF assigns class-specific roles and accumulates claim revisions that feed later layers (Sections 2–4).

2. Independent evidence for explanation quality (Critical Weakness 2; Question 2). We test explanations against generator-defined feature and interaction effects, in addition to replaying model scores. Across five synthetic worlds trained with class labels alone, QBAF's gradient-times-amplitude ranking reaches evidence average precision (AP) 0.726410.72641, against 0.516830.51683 for random and 0.531470.53147 for the native message ranking (Table 22). In two additional worlds, supervision on signed leaf-pair effects raises held-out interaction AP from 0.014430.01443 to 0.965940.96594 and from 0.018860.01886 to 0.948540.94854. Under identical pair supervision, direct-signed obtains 0.970410.97041 and 0.939780.93978, respectively (Tables 24–25). The reference is an independently defined effect on the generator's output.

3. Population-scale intervention tests (Critical Weakness 3). We extend the four traces to 600 fixed targets, 200 each on Cora, Photo, and Actor; 575 admit non-self relation edits. Deletion, fixed-normalization masking and retention are evaluated at matched budgets and two checkpoints. Controls enumerate all legal sets when there are at most 20 and sample 20 distinct sets otherwise. Bootstrap intervals resample targets, retaining their two checkpoints together (Table 19). Native relation selection exceeds random deletion by gap AUC 0.0070320.007032 on 175 Photo targets (95% CI [0.002715,0.011462][0.002715,0.011462]) and 0.0058160.005816 on 200 Actor targets ([0.004593,0.007036][0.004593,0.007036]). Holding the Actor QBAF predictor, target sets and budgets fixed, native ranking exceeds GNNExplainer in deletion by 0.0051520.005152 and masking by 0.0056470.005647, while retention gap falls by 0.0062650.006265 ([−0.009749,−0.003277][-0.009749,-0.003277]; Tables 3, 20). Across complete Actor predictors, all 12 pointwise paired intervals for deletion and masking favor QBAF over direct-signed, FAGCN and SES at two checkpoints, each with its own margin and ranking (Table 21).

4. Mechanism ablations and weighting (Critical Weakness 4; Question 3). Matched experiments isolate composition, receiver conditioning, role parameterization, claim feedback and class weighting. Removing atomic/joint local composition lowers accuracy on all four ablation graphs, by 1.071.07 points on Cora and 1.641.64 on Texas (Table 2). The matched direct-signed control retains the same arguments, composition, claims, readout and weighting: three roles gain 1.4781.478 Cora accuracy points under EM (95% CI [0.879,2.077][0.879,2.077]), and 1.9591.959 Actor macro-F1 points under unit weights ([0.726,3.192][0.726,3.192]), with test NLL lower by 0.08010.0801 (Table 16). Receiver conditioning is separately measured in Table 15; its balanced accuracy effects are +0.025,+0.025,+0.038+0.025,+0.025,+0.038 on Cora, Photo and Actor. Applying EM weighting to both QBAF and GPRGNN+JDR on Cora improves their accuracies by 3.5963.596 and 1.9381.938 points, respectively, an architecture-by-weighting interaction of +1.659+1.659 ([0.441,2.876][0.441,2.876]; Table 17).

5. Paired predictive improvements (Critical Weakness 5; Question 4). QBAF has the highest reported mean on six of nine benchmarks against 16 published baselines (Table 1). Table 12 adds ten-split, two-initialization matched comparisons with exact sign-flip tests and Holm correction. Five gains survive: Citeseer +1.623+1.623, Pubmed +0.773+0.773 and WikiCS +2.649+2.649 against Specformer; Photo +1.594+1.594 against NAGphormer; Physics +0.139+0.139 against SGFormer. The Physics paired 95% CI is [0.047,0.231][0.047,0.231] with Holm-adjusted p=0.0176p=0.0176. In separate shared-split comparisons, QBAF exceeds Specformer on Questions by 3.7203.720 ROC-AUC points ([2.675,4.765][2.675,4.765]; Table 11) and GATv2 on Actor by 6.1696.169 accuracy points ([4.889,7.448][4.889,7.448]; Appendix C.3). On Questions, validation-selected unit weighting raises QBAF by 0.9510.951 ROC-AUC points; its paired difference from GAT becomes −0.057-0.057 ([−0.732,0.617][-0.732,0.617]; Table 14). On Texas, terminal claims improve accuracy over initial claims by 10.16410.164 points (Table 18).

Thank you for the response. I’ve updated my review to reflect your clarifications and the additional results, but I’ve decided to keep my original score.

AuthorsOpenReview ↗

Thank You Very Much for Your Constructive Feedback

Thank you very much for taking the time to revisit our manuscript and consider the additional results. We sincerely appreciate your detailed and constructive feedback, which has substantially strengthened the paper.

Your comments motivated several important additions, including a matched comparison between three-role and direct-signed propagation (Table 16), independent synthetic ground-truth evaluations of explanation quality (Tables 22–25), and paired statistical analyses with Holm correction (Table 12).

In particular, the original four illustrative traces have now been extended to a systematic intervention study covering 600 targets, including 575 with editable non-self relations. Section 5.4 and Tables 19–21 provide quantitative deletion, masking, and retention evaluations, together with comparisons against established explanation methods.

These experiments have helped us develop a clearer understanding of the role-dependent intervention mechanism and strengthen the empirical foundations of QBAF. We are grateful for the care and expertise you have brought to the review process and for the valuable improvements your feedback has inspired.

Reviewer_k1j3OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper combines Quantitative Bipolar Argumentation Framework (QBAF) with node classification. Classification can be reconstructed through the graph structure. Instead of treating graph connections simply as relationships between nodes, QBAF models them as supporting, attacking, or neutral relationships between arguments associated with different classes. It combines node features and class-specific arguments to calculate prediction scores, which makes them traceable throughout the graph structure.

Critical strengths

Integrating structured argumentation into graph-based classification while making predictions more traceable and open to analysis.

Critical weaknesses

The paper introduces an extensive mathematical formulation of QBAF, but the connection to foundational work on quantitative bipolar argumentation and the specific semantics adopted should be made more explicit. The authors should clearly distinguish established formalism from their proposed vector-valued extension and provide direct references at the relevant definitions.

Furthermore, although Table 1 reports the sample standard deviation over ten runs, the discussion focuses primarily on the mean performance and does not sufficiently address the high variability observed on some datasets.

The paper does not explicitly discuss its limitations.

Questions for authors
  • What are the main limitations of QBAF in terms of scalability, sensitivity to architectural choices, and generalization beyond the nine evaluated datasets?
  • On some datasets, the performance gains are small in comparison to the standard deviation of QBAF's across runs. Could you please elaborate on the high values?
  • How did you choose the base architecture?
4 comments
AuthorsOpenReview ↗

QBAF Semantics, Statistical Evidence, Architecture Design, and Scalability

1. Foundations and adopted semantics (Critical Weakness 1). Section 2 explicitly distinguishes classical scalar QBAFs from our learned finite-depth vector-valued semantics. It cites Dung (1995) for attacks, Cayrol and Lagasquie-Schiex (2005) for bipolar support, and Baroni et al. (2018, Definitions 1–2) for scalar quantitative frameworks, base scores and evaluated strengths. The semantics adopted is Semθ,LV(G,X)=(Tθ,L∘⋯∘Tθ,1)(Eθ(X)).\mathrm{Sem}^{V}_{\theta,L}(G,X)=(T_{\theta,L}\circ\cdots\circ T_{\theta,1})(E_{\theta}(X)).

Equation (1) defines this operator, and Eqs. (2)–(7) implement its updates. Each node maintains a matrix of class-specific argument evidence and a vector of accumulated raw class claims. Formally, Hiℓ∈RK×CH_i^\ell\in\mathbb R^{K\times C} and qiℓ∈RCq_i^\ell\in\mathbb R^C. Every layer composes atomic and joint evidence, propagates messages under support, neutral and attack roles, and revises the claims. With fitted parameters fixed, evaluation terminates after exactly L layers, including on cyclic graphs. Appendix A.5 proves that a single class's states and raw scores are reproduced exactly by its parameter slice. Appendix A.6 compares DF-QuAD (Rago et al., 2016) and MLP-based semantics (Potyka, 2021) on common examples and contrasts their update rules with ours (Tables 6–7). Appendix A.7 independently verifies the layer equations and class projection in 48 cases.

2. Run variability and paired statistical evidence (Critical Weakness 2; Question 2). Table 1 reports sample SD across ten QBAF fits: two initializations on each of five splits. Table 13 decomposes the observed variation: differences between splits account for 94.1% on Cora and 91.5% on Citeseer; differences between runs on the same split account for 77.0% on Questions. Texas has a 4.03-point sample SD; between-split and within-split differences account for 54.0% and 46.0% of the total sum of squares, respectively. Each Texas test split has 61 nodes, so one additional correct prediction changes accuracy by 1.639 points.

A model's across-run SD measures fitted-score variation, whereas a matched comparison measures method differences on the same splits. Table 12 averages two initializations within each of ten common splits, applies exact paired sign-flip tests and corrects nine comparisons using Holm's method. Five gains survive: Citeseer +1.623, Pubmed +0.773 and WikiCS +2.649 points versus Specformer; Photo +1.594 versus NAGphormer; and Physics +0.139 versus SGFormer. Physics has a paired 95% CI [0.047, 0.231] and Holm-adjusted p=0.0176. On Cora, QBAF gains +0.764 points over GPRGNN+JDR ([0.006, 1.521], Holm-adjusted p=0.129). Appendix C.1 distinguishes published baseline SDs from confidence intervals, including GPRGNN+JDR's bootstrap interval on Cora. The Cora ±0.25 is a 95% bootstrap CI half-width.

3. Base architecture and component evidence (Question 3). The 64 argument slots form a six-bit index: each slot reads six XOR neighbors for atomic and joint local evidence composition. Of the 64 slots, 56 carry learned feature predicates and eight preserve class-specific source projections from four feature groups. Four graph heads each share a role assignment across 16 distinct payload slots, reducing role evaluations while retaining individual evidence coordinates. Roles use source and receiver contexts; accumulated claims feed subsequent layers and enable exact class-specific traces. In matched retraining, removing local composition reduces mean accuracy on all four tested graphs, including 1.07 points on Cora and 1.64 on Texas (Table 2). Tables 15–16 separately examine receiver conditioning, class weighting and three-role versus scalar-gate parameterization. Width, heads and partner count were fixed. Depth, learning rate, weight decay and dropout were selected through validation alone. The original benchmark used ten layers except Texas (four); matched Specformer reruns selected two QBAF layers on Photo and Actor.

4. Limitations, scalability and generalization (Critical Weakness 3; Question 1). Section 5.5 explicitly discusses computational cost, architectural sensitivity and generalization. One dense FP32 argument state requires 4NKC bytes; class blocking and sparse propagation reduce transient storage and preserve reference predictions. On the official ogbn-arxiv temporal split (169,343 nodes, 40 classes), ten-layer QBAF achieves 71.53 ± 0.31% test accuracy across five initializations, 2.73 points above GPR-GNN. On an RTX 5090 D, a full-graph update takes 2.150 s, inference 0.608 s, and peak allocated training/inference memory is 23.840/3.236 GiB (Tables 4, 32–33). This measured computation supports explicit class-specific evidence, role evaluations, cumulative claim revisions and executable edits. Independent width/head sensitivity has not been measured. All ten evaluated tasks are transductive; generalization to unseen graphs has not been evaluated.

Reviewer_k1j3OpenReview ↗

Answer to QBAF Semantics, Statistical Evidence, Architecture Design, and Scalability

Dear authors, thank you very much for your response. However, I am a little confused as I am having some difficulty matching your response to my concerns. I explicitly asked about the limitations of your work, the high variability of the results, and the base architecture. In your response, you address weaknesses and questions that I did not ask. Have you perhaps mixed something up?

AuthorsOpenReview ↗

Correction: QBAF Semantics, Statistical Variability, Architecture, and Limitations

We sincerely apologize for the mix-up. Our earlier response under your review inadvertently contained text prepared for a different reviewer. We deeply regret the confusion and the unnecessary time this has cost you, and we appreciate your patience in bringing this to our attention. Your thoughtful and constructive feedback has genuinely helped us make substantial improvements to the manuscript, for which we are sincerely grateful. We have now carefully aligned the response below with your three Critical Weaknesses and three Questions for Authors.

AuthorsOpenReview ↗

Corrected Response to Reviewer k1j3: Semantics, Variability, Architecture, and Limitations

1. Foundations and adopted semantics (Critical Weakness 1). Section 2 explicitly distinguishes classical scalar QBAFs from our learned finite-depth vector-valued semantics. It cites Dung (1995) for attacks, Cayrol and Lagasquie-Schiex (2005) for bipolar support, and Baroni et al. (2018, Definitions 1–2) for scalar quantitative frameworks, base scores and evaluated strengths. The semantics adopted is Semθ,LV(G,X)=(Tθ,L∘⋯∘Tθ,1)(Eθ(X)).\mathrm{Sem}^{V}_{\theta,L}(G,X)=(T_{\theta,L}\circ\cdots\circ T_{\theta,1})(E_{\theta}(X)).

Equation (1) defines this operator, and Eqs. (2)–(7) implement its updates. Each node maintains a matrix of class-specific argument evidence and a vector of accumulated raw class claims. Formally, Hiℓ∈RK×CH_i^\ell\in\mathbb R^{K\times C} and qiℓ∈RCq_i^\ell\in\mathbb R^C. Every layer composes atomic and joint evidence, propagates messages under support, neutral and attack roles, and revises the claims. With fitted parameters fixed, evaluation terminates after exactly L layers, including on cyclic graphs. Appendix A.5 proves that a single class's states and raw scores are reproduced exactly by its parameter slice. Appendix A.6 compares DF-QuAD (Rago et al., 2016) and MLP-based semantics (Potyka, 2021) on common examples and contrasts their update rules with ours (Tables 6–7). Appendix A.7 independently verifies the layer equations and class projection in 48 cases.

2. Run variability and paired statistical evidence (Critical Weakness 2; Question 2). Table 1 reports sample SD across ten QBAF fits: two initializations on each of five splits. Table 13 decomposes the observed variation: differences between splits account for 94.1% on Cora and 91.5% on Citeseer; differences between runs on the same split account for 77.0% on Questions. Texas has a 4.03-point sample SD; between-split and within-split differences account for 54.0% and 46.0% of the total sum of squares, respectively. Each Texas test split has 61 nodes, so one additional correct prediction changes accuracy by 1.639 points.

A model's across-run SD measures fitted-score variation, whereas a matched comparison measures method differences on the same splits. Table 12 averages two initializations within each of ten common splits, applies exact paired sign-flip tests and corrects nine comparisons using Holm's method. Five gains survive: Citeseer +1.623, Pubmed +0.773 and WikiCS +2.649 points versus Specformer; Photo +1.594 versus NAGphormer; and Physics +0.139 versus SGFormer. Physics has a paired 95% CI [0.047, 0.231] and Holm-adjusted p=0.0176. On Cora, QBAF gains +0.764 points over GPRGNN+JDR ([0.006, 1.521], Holm-adjusted p=0.129). Appendix C.1 distinguishes published baseline SDs from confidence intervals, including GPRGNN+JDR's bootstrap interval on Cora. The Cora ±0.25 is a 95% bootstrap CI half-width.

3. Base architecture and component evidence (Question 3). The 64 argument slots form a six-bit index: each slot reads six XOR neighbors for atomic and joint local evidence composition. Of the 64 slots, 56 carry learned feature predicates and eight preserve class-specific source projections from four feature groups. Four graph heads each share a role assignment across 16 distinct payload slots, reducing role evaluations while retaining individual evidence coordinates. Roles use source and receiver contexts; accumulated claims feed subsequent layers and enable exact class-specific traces. In matched retraining, removing local composition reduces mean accuracy on all four tested graphs, including 1.07 points on Cora and 1.64 on Texas (Table 2). Tables 15–16 separately examine receiver conditioning, class weighting and three-role versus scalar-gate parameterization. Width, heads and partner count were fixed. Depth, learning rate, weight decay and dropout were selected through validation alone. The original benchmark used ten layers except Texas (four); matched Specformer reruns selected two QBAF layers on Photo and Actor.

4. Limitations, scalability and generalization (Critical Weakness 3; Question 1). Section 5.5 explicitly discusses computational cost, architectural sensitivity and generalization. One dense FP32 argument state requires 4NKC bytes; class blocking and sparse propagation reduce transient storage and preserve reference predictions. On the official ogbn-arxiv temporal split (169,343 nodes, 40 classes), ten-layer QBAF achieves 71.53 ± 0.31% test accuracy across five initializations, 2.73 points above GPR-GNN. On an RTX 5090 D, a full-graph update takes 2.150 s, inference 0.608 s, and peak allocated training/inference memory is 23.840/3.236 GiB (Tables 4, 32–33). This measured computation supports explicit class-specific evidence, role evaluations, cumulative claim revisions and executable edits. Independent width/head sensitivity has not been measured. All ten evaluated tasks are transductive; generalization to unseen graphs has not been evaluated.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.