- Rating
- 3(+1)
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
MAGIC does replay-free GFSCIL with a frozen backbone (SGC, S²GC, PPR, heat kernel, or a frozen pretrained encoder) and a ridge classifier updated in closed form from accumulated sufficient statistics. Two additions- (1) A topological prior: edges in the base graph are binned by cosine similarity, each bin gets a signed log-ratio of same-class to different-class edge mass, and these become Potts potentials on a k-hop subgraph around the supports, with mean-field posteriors projected onto novel classes and folded back into the statistics. (2) drift-aware distillation: cross-session edges move historical representations even with a frozen encoder, so the old classifier’s softened predictions are reimposed on the new representations, weighted by drift. Two theorems cover the perturbation identity and mean-field contraction. Five datasets, eight baselines, +5.48 MA / +9.33 FA and −10.78 PD at 5-shot, with much lower training time.
Critical strengths
(1) Results are strong and robust across conditions. Coauthor CS is roughly +12 MA and +20 FA over the best baseline, and the gains survive at 1- and 3-shot (Tables 6, 7) and across five backbones (Table 2). The backbone study is the most convincing part: it rules out dependence on one feature extractor.
(2) The supervision-mass abstraction is clean. Hard labels, distillation targets and MRF pseudo-labels all enter the same R and C statistics (Eqs. 2, 7, 13), derived properly in A.1.6. Removing gradient training from incremental sessions is a real practical win, and Figure 3 follows from that rather than from tuning.
Critical weaknesses
(1) The gains aren’t attributed. MAGIC differs from every baseline in three ways, frozen features, closed-form class-balanced ridge, and the two modules, and only the third is ablated. Misiing ablation is frozen SGC plus class-balanced ridge, no prior, no distillation. Table 3 removes one component at a time, but not both.
This distiction is necessary because on ogbn-arxiv w/o Prior gets 53.69 MA / 37.29 FA, below TAP at 57.97 / 42.69, so the whole advantage there is prior injection. But Figure 6(a) is flat in η from 1 to 6, while 6(b) climbs steeply in ρ, and ρ = 0 on CoraFull is around 59–60 MA, roughly TAP. That’s consistent with the benefit coming from injecting pseudo-supervision mass at all, i.e. transductive self-training, with the topology-derived coupling adding little. The experiment that would rule this out, η = 0 or constant coupling, isn’t there.
Same area, protocol question: MRF refinement pseudo-labels the unlabeled k-hop neighborhood of the supports. If evaluation nodes sit in U_t, MAGIC is self-training on the test set, which inductive TAP cannot do. The paper doesn’t say whether they’re disjoint.
(2) Both contributions are asserted, and the theory doesn’t close this gap. The prior’s headline property is that signed β_b handles heterophily, but every dataset has h_edge between 0.57 and 0.81 (Table 4). No heterophilous graph is tested, so I can’t tell whether negative couplings are ever recovered or whether this reduces to label propagation here.
Distillation reports nothing about A_t, the set it targets: not its size, not the spread of δ_i, not accuracy on those nodes with and without. The aggregate effect is +1.6 MA / +1.8 FA on ogbn-arxiv, small enough that uniform-mass distillation might match it, leaving the drift weighting in Eq. 5 unmotivated.
Theorem 1 expands (H⁻+Δ)ᵀD(H⁻+Δ) and shows the statistics change, which was never in doubt; it says nothing about forgetting or about this correction. Theorem 2’s condition η‖J_FF‖₂/2 < 1 is never checked at the settings used, where η reaches 6 (Table 11).
Finally, computing δ_i and the teacher targets requires keeping h_{t-1,i} for every historical node plus the old structure. That’s O(|V|·d) stored representations, a form of representation replay, which doesn't make sense with replay-free.
Questions for authors
Questions for Authors
(1) Could the authors report the performance of frozen SGC with class-balanced ridge regression, without either the topology prior or distillation, across all five datasets? Table 3 only removes one component at a time. This joint ablation would help establish how much of the improvement comes from the proposed modules rather than the analytic classifier itself.
(2) In Figure 6, performance barely changes with , whereas increasing gives noticeable gains. Could the authors report results with (keeping fixed) and with constant coupling ( across all bins)? This would clarify whether the gains come from the learned topological relationships or simply from adding pseudo-supervision.
(3) Are evaluation nodes included in the MRF candidate set ? If so, how is this accounted for when comparing against inductive methods such as TAP? It would be useful to see the performance when evaluation nodes are excluded from pseudo-label refinement.
(4) Could the authors provide more analysis of the affected historical nodes ? In particular, what fraction of historical nodes are affected, how much do their representations drift, and how does distillation change accuracy specifically on these nodes? A comparison with uniform distillation weights would also help justify the drift-based weighting in Eq. (5).
(5) The paper claims that the signed topological prior handles heterophilous relationships, but all five datasets have edge homophily above 0.57 (Table 4). Could the authors evaluate on a heterophilous dataset such as Chameleon, Squirrel, or Actor and show the learned values? This would help establish whether the method actually learns and benefits from negative relational couplings.
(6) Some relevant works appear to be missing from the discussion: GOTHAM (TMLR 2025), which studies graph class-incremental learning under weak supervision using topology-based prototypes and distillation; G-Scaler (Neurocomputing 2026), which addresses structural transfer and representation drift in GFSCIL; and Meta-Tuner (ESWA 2025), which uses node-specific transformations and prototype regularization. Could the authors clarify how MAGIC differs from these methods and provide comparisons where the experimental settings are compatible?
6 comments
Weakness 2 & Q4 & Q5 & Q6
Weakness 2
Relational Prior, Distillation Mechanism, Theoretical Scope, and Memory
Weakness 2.1 & Q5 Relational Prior on Heterophilous Relations
- Inter-Class Edges in Homophilous Graphs: Even in graphs with higher overall homophily, inter-class edges are prevalent, homophily assumptions fail on these edges.
- Limitation on Purely Heterophilous Graphs: Parameter-free backbones and existing pretrained Graph Foundation Models struggle to extract high-quality node features on fully heterophilous graphs. Consequently, applying MAGIC to purely heterophilous graphs remains challenging. We will tone down this over-claim in the final revision.
Weakness 2.2 & Q4 Drift Distillation
We evaluated the final-session classification accuracy on the affected historical nodes on ogbn-arxiv (5-shot, SGC backbone):
| Distillation Variant | Final Accuracy on Affected Nodes (%) |
|---|---|
| w/o Distillation | 45.30 ± 1.39 |
| Drift-Weighted Distillation | 48.60 ± 2.93 |
| Uniform-Mass Distillation | 50.60 ± 2.88 |
- Effectiveness of Distillation: Distillation provides a substantial performance gain on affected historical nodes.
- Design Refinement: Our original design aimed to allocate more mass to nodes with more severe drift, but experiments show that simple uniform mass allocation works better. We will update the method accordingly in the final version.
In the final version, we will provide more detailed experiments on the distillation mechanism.
Weakness 2.3 Theoretical Role and Convergence Conditions (Theorems 1 & 2)
- Theorem 1 (Perturbation Formulation): Theorem 1 demonstrates that fixed encoder parameters alone do not guarantee invariant historical statistics on dynamic graphs.
- Theorem 2 (Mean-Field Convergence): The condition is a standard sufficiency condition under the Banach fixed-point theorem that guarantees convergence to a unique fixed point. Actually, even when this condition is not satisfied, damped mean-field updates still converge to a stationary fixed point. We will revise Theorem 2 accordingly in the final version.
Weakness 2.4 Transductive Setting and Memory Overhead
Accessing historical nodes is fully legitimate as MAGIC strictly follows the standard transductive GFSCIL setting. Without accessing these nodes, it is extremely difficult to correct the representation drift caused by cross-session edges. Moreover, accessing only the affected nodes already incurs very small overhead.
Q6: Explanation of baselines
Table R1: Comparison with recent GFSCIL methods on Coauthor CS (5-shot, %)
| Method | Mean Acc (MA) | Final Acc (FA) | Mean Novel (AN) | Final Novel (FN) | PD () |
|---|---|---|---|---|---|
| G-Scaler | 68.94 ± 4.24 | 44.89 ± 11.63 | 51.10 ± 6.19 | 42.12 ± 9.71 | 46.30 ± 12.86 |
| Meta-Tuner | 85.22 ± 2.79 | 73.71 ± 3.45 | 58.12 ± 6.49 | 49.78 ± 7.43 | 23.33 ± 3.93 |
| MAGIC (SGC) | 94.03 ± 0.79 | 91.45 ± 0.28 | 92.02 ± 1.70 | 90.87 ± 0.70 | 5.54 ± 1.34 |
MAGIC substantially outperforms our reproductions of G-Scaler and Meta-Tuner across all metrics. We summarize the key methodological differences between our work and G-Scaler and Meta-Tuner:
- G-Scaler (Neurocomputing 2026): Addresses few-shot session overfitting purely in the feature space and primarily addresses drift caused by continual GNN training via a hybrid-space projector. In contrast, MAGIC is gradient- and replay-free via closed-form updates on a frozen backbone, mitigates few-shot overfitting by injecting graph topology directly into label inference via a signed Potts MRF prior over unlabeled -hop neighborhoods, and explicitly corrects topology-induced drift from cross-session edges via analytic distillation.
- Meta-Tuner (ESWA 2025): Meta-trains a node-conditioned hypernetwork with simplex prototype regularization and relies solely on labeled support prototypes at inference. MAGIC requires no meta-training, leverages unlabeled neighborhood topology to mitigate few-shot overfitting, and explicitly compensates for cross-session historical node drift.
I thank the authors for doing rigorous experimentation along different axes but most of the concerns remain unresolved, I am open to changing my score, provided the concerns are fully addressed here and not left for the updation in the manuscript.
Major remaining concerns
(1) In W1/Q1 the ablation is essential to undersatnd what each components contributes, without that the experiments are kind of ad-hoc. (2) Again in W1/Q2 η = 0 or constant coupling, isn’t there. So the experiments are still not convincing enough to isolate the method. (3) the drift distillation method W2.2 and Q4- gives a opposing view to the method in the paper, its good to acknowledge that but the original drift-dependent weighting motivation seems to have gone.
Weakness 1 & Q1 & Q2 & Q3
Weakness 1
Weakness 1.1 & Q1 Full Joint Ablation Baseline
In the final version, we will include the complete joint ablation baseline across all five datasets.
Weakness 1.2 & Q2 Disentangling Topological Coupling (η) from Pseudo-Label Mass (ρ)
Table R2: Performance across varying relation strengths η on ogbn-arxiv (%)
Table R3: Performance across varying relation strengths η on CoraFull (%)
Therefore, the prior injection strength η has a significant impact on model performance.
Therefore, the results observed in the hyperparameter experiments are not contradictory.
Weakness 1.3 & Q3 Clarification on the Transductive Protocol and Test Set Isolation
As shown above, the performance difference between strictly excluding test nodes and the standard setup is negligible.