acceptodds
Under review as a conference paper at ICLR 2027

MAGIC: Topology-Aware Analytic Graph Few-Shot Class-Incremental Learning

17 views

Abstract

Graph few-shot class-incremental learning (GFSCIL) requires a model to continually recognize emerging classes from only a few labeled nodes while preserving previously acquired knowledge. Beyond the catastrophic forgetting inherited from conventional graph continual learning, GFSCIL presents two distinctive challenges: extremely limited novel-class supervision causes severe overfitting, while cross-session edges—edges connecting newly arriving nodes with historical nodes—alter historical propagation neighborhoods and thereby induce representation drift. We propose MAGIC, a replay-free GFSCIL framework that combines a frozen graph representation backbone (e.g., an intrinsically parameter-free backbone such as SGC or a pretrained graph foundation model) with closed-form analytic continual learning. To alleviate novel-session overfitting, MAGIC learns a topological prior from the base graph that can characterize both homophilous and heterophilous relations, and injects this prior through Potts Markov random field inference to refine supervision for novel classes. To mitigate representation drift, MAGIC transfers previous predictions from the old representations of affected historical nodes to their updated representations through drift-aware analytic distillation. Experiments across five datasets and eight baselines demonstrate the effectiveness of MAGIC. Under the 5-shot setting, MAGIC improves Mean Accuracy and Final Accuracy by 5.48% and 9.33% on average, and reduces Performance Drop by 10.78% on average compared with the best baselines. MAGIC also shows clear advantages under the 1- and 3-shot settings, with larger gains as the number of supports increases. Moreover, MAGIC requires substantially less training time.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph Few-Shot Class-Incremental Learning, Graph Continual Learning, Graph Neural Networks, Analytic Continual Learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_B66wOpenReview ↗
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 η\eta, whereas increasing ρ\rho gives noticeable gains. Could the authors report results with η=0\eta=0 (keeping ρ\rho fixed) and with constant coupling (βb=β\beta_b=\beta 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 UtU_t? 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 AtA_t? 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 βb\beta_b 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
AuthorsOpenReview ↗

Weakness 1 & Q1 & Q2 & Q3

Weakness 1

Gain Attribution, Prior vs. Mass, and Transductive Setting

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\eta) from Pseudo-Label Mass (ρ\rho)

  • Significance of the Topological Prior: To evaluate the topological prior, we report results across η∈{0,1,2,3,4,5,6}\eta \in \{0, 1, 2, 3, 4, 5, 6\} on ogbn-arxiv and CoraFull (5-shot, SGC backbone, ρ=1\rho=1):

Table R2: Performance across varying relation strengths η\eta on ogbn-arxiv (%)

η\etaMean AccFinal AccMean NovelFinal NovelPD
062.14 ± 4.8947.22 ± 3.3043.74 ± 7.8132.57 ± 6.3336.54 ± 2.89
162.77 ± 3.7149.13 ± 2.2647.51 ± 4.5936.67 ± 5.7835.05 ± 2.72
263.97 ± 3.9750.34 ± 2.3849.63 ± 5.2038.29 ± 5.0933.84 ± 2.39
364.49 ± 3.9350.90 ± 2.5250.64 ± 5.2839.19 ± 4.7933.28 ± 2.39
464.75 ± 3.8851.13 ± 2.5551.16 ± 5.3939.60 ± 4.7033.05 ± 2.41
564.93 ± 3.8151.31 ± 2.4851.51 ± 5.4739.95 ± 4.7332.87 ± 2.40
665.04 ± 3.8051.37 ± 2.4751.73 ± 5.5340.07 ± 4.7132.81 ± 2.42

Table R3: Performance across varying relation strengths η\eta on CoraFull (%)

η\etaMean AccFinal AccMean NovelFinal NovelPD
063.93 ± 0.7153.71 ± 1.1952.47 ± 1.7047.70 ± 2.1726.59 ± 3.58
165.07 ± 0.9955.58 ± 1.0256.31 ± 1.5451.36 ± 2.3024.63 ± 3.54
265.28 ± 0.9955.86 ± 1.0556.47 ± 1.8651.65 ± 2.1924.35 ± 3.56
365.46 ± 0.9556.04 ± 1.1056.58 ± 2.2951.83 ± 2.2324.17 ± 3.68
465.50 ± 0.9356.18 ± 1.1256.48 ± 2.6651.93 ± 2.3624.02 ± 3.71
565.52 ± 0.8756.37 ± 1.1656.37 ± 2.8652.11 ± 2.3523.83 ± 3.75
665.53 ± 0.8656.39 ± 1.0756.24 ± 3.0152.08 ± 2.2423.82 ± 3.68
  1. Decisive Empirical Gains over η=0\eta = 0: Compared to η=0\eta = 0, enabling MRF prior injection brings marked improvements across all metrics.
  2. Growth Trend: On CoraFull, accuracy jumps sharply at η=1\eta = 1 and largely saturates beyond η≥2\eta \ge 2; on ogbn-arxiv, a similarly strong rise occurs for η≤2\eta \le 2 before plateauing.

Therefore, the prior injection strength η\eta has a significant impact on model performance.

  • Roles of η\eta and ρ\rho: There is no inherent correlation between the model's behavioral trends when varying η\eta versus varying ρ\rho:
    1. η\eta controls the strength of relational evidence. As demonstrated by the empirical data, once η\eta reaches a certain threshold, its subsequent impact on pseudo-label quality becomes relatively small.
    2. ρ\rho acts as the supervision budget. When the pseudo-labels refined by the MRF are of high quality, scaling up ρ\rho can improve the performance gain.

Therefore, the results observed in the hyperparameter experiments are not contradictory.

Weakness 1.3 & Q3 Clarification on the Transductive Protocol and Test Set Isolation

  • Standard Transductive Formulation: Following the standard GFSCIL formulation, the setting is transductive: the visible graph contains both labeled and unlabeled nodes observed up to session tt. Unlabeled nodes and topological edges are accessible for feature propagation and semi-supervised inference, but all ground-truth evaluation labels are strictly withheld. No evaluation labels are ever exposed.
  • Empirical Validation without Test Nodes in UtU_t: To address the concern of potential self-training bias on test nodes, we conducted an experiment strictly excluding all test/evaluation nodes from the kk-hop candidate set UtU_t on ogbn-arxiv (η=6\eta = 6, 5-shot):
SettingMean AccFinal AccMean NovelFinal NovelPD
w/o test64.88 ± 3.8451.24 ± 2.3751.33 ± 5.5739.84 ± 4.9232.94 ± 2.55
ordinary65.04 ± 3.7851.37 ± 2.4551.72 ± 5.5140.09 ± 4.7332.81 ± 2.42

As shown above, the performance difference between strictly excluding test nodes and the standard setup is negligible.

AuthorsOpenReview ↗

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 VariantFinal Accuracy on Affected Nodes At\mathcal{A}_t (%)
w/o Distillation45.30 ± 1.39
Drift-Weighted Distillation48.60 ± 2.93
Uniform-Mass Distillation50.60 ± 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 η∥JFF∥2/2<1\eta \|J_{FF}\|_2 / 2 < 1 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, %)

MethodMean Acc (MA)Final Acc (FA)Mean Novel (AN)Final Novel (FN)PD (↓\downarrow)
G-Scaler68.94 ± 4.2444.89 ± 11.6351.10 ± 6.1942.12 ± 9.7146.30 ± 12.86
Meta-Tuner85.22 ± 2.7973.71 ± 3.4558.12 ± 6.4949.78 ± 7.4323.33 ± 3.93
MAGIC (SGC)94.03 ± 0.7991.45 ± 0.2892.02 ± 1.7090.87 ± 0.705.54 ± 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 kk-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.

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

This paper proposes MAGIC, a replay-free framework for graph few-shot class-incremental learning that pairs a frozen graph backbone with closed-form analytic continual learning. MAGIC counters cross-session representation drift by distilling previous predictions onto the updated representations of affected historical nodes. It alleviates novel-class overfitting by injecting a base-learned relational prior through Potts MRF inference to produce pseudo-supervision.

Critical strengths

1.Efficient and replay-free design. The closed-form analytic updates reduce training time to seconds while keeping accuracy competitive, which I consider the main practical merit of the work.

2.Recognition of cross-session representation drift. The authors note that a frozen encoder does not keep historical representations fixed when new edges arrive, a reasonable observation largely overlooked in prior GFSCIL studies.

Critical weaknesses

1.Insufficient comparison with recent work. Several recent GFSCIL methods are neither compared nor discussed, including MEGA (2025), URPD (Multimedia Systems, 2025), and GOTHAM (TMLR 2025). URPD, like MAGIC, also exploits unlabeled nodes to counter overfitting. Moreover, MAGIC does not surpass results already reported in the literature. For example, Mecoin reports 70.90 mean and 61.36 final accuracy on CoraFull, both higher than MAGIC’s 65.45 and 56.02, although under a different split, while its reproduction in this paper reaches only 47.57. Without evaluation under established protocols, the state-of-the-art claim is not convincing.

2.The contribution of the topological prior is not substantiated. Performance is nearly unchanged when the number of similarity bins drops to one (Fig. 6d), which reduces the prior to uniform label propagation. The pseudo-label mass ρ dominates the sensitivity results, and the claimed ability to handle heterophily is untested, since all five datasets are homophilous.

Questions for authors

1.Can the authors evaluate MAGIC under the original protocols of existing methods (e.g., Mecoin’s or TAP’s splits on CoraFull)? Can they also compare against recent GFSCIL methods such as MEGA, URPD, and GOTHAM? This would show whether the claimed gains hold beyond the authors’ own setting.

2.The reproduced baselines are notably lower than their published results (e.g., Mecoin: 47.57 here vs. 70.90 reported). How were the baselines implemented and tuned, and what explains this gap?

3.How much of the improvement comes from the similarity-dependent prior itself? Please report two variants: (a) pseudo-labels from the naive classifier without MRF refinement (η = 0), and (b) the MRF with a single constant β, i.e., B = 1, on all datasets.

7 comments
AuthorsOpenReview ↗

Weakness 1 & Q1 & Q2

Weakness 1

Weakness 1.1 & Q1: Baseline Comparisons

  • Comparison with Recent GFSCIL Works (MEGA, URPD, G-Scaler, Meta-Tuner): Because none of MEGA, URPD, G-Scaler, or Meta-Tuner have publicly released source code, we re-implemented these methods based on their respective papers and evaluated them under our unified 5-shot GFSCIL protocol on Coauthor CS. As shown in Table R1, MAGIC achieves the best performance across all metrics,.

Table R1: Comparison with recent GFSCIL methods on Coauthor CS (5-shot, %)

MethodMean Acc (MA)Final Acc (FA)Mean Novel (AN)Final Novel (FN)PD (↓\downarrow)
URPD31.79 ± 1.9512.86 ± 3.1243.02 ± 7.3531.42 ± 11.6077.36 ± 2.72
G-Scaler68.94 ± 4.2444.89 ± 11.6351.10 ± 6.1942.12 ± 9.7146.30 ± 12.86
MEGA78.40 ± 8.2164.26 ± 12.0252.23 ± 25.0039.24 ± 15.1730.01 ± 12.81
Meta-Tuner85.22 ± 2.7973.71 ± 3.4558.12 ± 6.4949.78 ± 7.4323.33 ± 3.93
MAGIC (SGC)94.03 ± 0.7991.45 ± 0.2892.02 ± 1.7090.87 ± 0.705.54 ± 1.34
  • Advantages of MAGIC over These Baselines:

    • vs. MEGA, Meta-Tuner, and URPD: These methods rely on meta-learning to alleviate novel-session overfitting, which incurs high computational overhead, easily overfits when base-session data are limited, and operates primarily in the node-feature space without exploiting graph topology. In contrast, MAGIC statistically computes a topological prior, making it both faster and more robust.
    • vs. G-Scaler: G-Scaler relies on labeled MM-query nodes in each novel session (LNCE\mathcal{L}_{NCE}) and accesses previous sessions' MM-query nodes (QtoldQ_t^{old}) in the latest session (LREC\mathcal{L}_{REC}), which constitutes label/sample replay. MAGIC is strictly replay-free and never accesses extra labeled nodes beyond the KK supports per novel class.
    • vs. All Four Baselines: None of them account for historical node representation drift caused by cross-session edges. MAGIC explicitly identifies affected historical nodes and preserves their predictions via drift-aware analytic distillation.
  • Scope of GOTHAM: GOTHAM (TMLR 2025) investigates weakly supervised graph class-incremental learning involving zero-shot novel classes and auxiliary semantic/text embeddings on text-attributed graphs. This problem setting relies on multimodal side information and zero-shot assumptions absent in standard GFSCIL, making direct comparison inapplicable.

Weakness 1.2 & Q2: Discrepancy with Published Numbers

  • Split Difficulty and Mecoin Discrepancy: The numerical difference arises from class split configurations rather than improper tuning. On CoraFull (70 classes total):
    • Our setting: 20 base classes and 50 novel classes, 5-way 1-, 3-, 5-shot.
    • Mecoin setting: 50 base classes and 20 novel classes, 2-way 5-shot.

To provide a direct comparison, we evaluated MAGIC under Mecoin's 50-base / 20-novel split on CoraFull (5-shot):

MethodBackboneS0 (Base)S1S2S3S4S5S6S7S8S9S10 (FA)PDMean Acc
Mecoin (published)GCN82.1881.5677.4771.5370.4369.7068.7868.0766.7062.1061.3620.8270.90
Mecoin (published)GAT75.5373.0470.7467.5266.0564.9764.1863.6862.0260.6260.1015.4366.22
MAGIC (Ours)SGC70.0669.6369.0368.0367.0366.0865.5864.7164.3463.4663.286.7866.48

Despite starting with a lower base accuracy (because Mecoin explicitly trains parameterized GNNs on the base session, leading to higher initial performance), MAGIC achieves a higher Final Accuracy and a dramatically lower Performance Drop.

AuthorsOpenReview ↗

Weakness 2 & Q3

Weakness 2 & Q3

Contribution of Topological Prior and Heterophily Evaluation

To evaluate the topological prior, we report results across η∈{0,1,2,3,4,5,6}\eta \in \{0, 1, 2, 3, 4, 5, 6\} on ogbn-arxiv and CoraFull (5-shot, SGC backbone, ρ=1\rho=1):

Table R2: Performance across varying relation strengths η\eta on ogbn-arxiv (%)

η\etaMean AccFinal AccMean NovelFinal NovelPD
062.14 ± 4.8947.22 ± 3.3043.74 ± 7.8132.57 ± 6.3336.54 ± 2.89
162.77 ± 3.7149.13 ± 2.2647.51 ± 4.5936.67 ± 5.7835.05 ± 2.72
263.97 ± 3.9750.34 ± 2.3849.63 ± 5.2038.29 ± 5.0933.84 ± 2.39
364.49 ± 3.9350.90 ± 2.5250.64 ± 5.2839.19 ± 4.7933.28 ± 2.39
464.75 ± 3.8851.13 ± 2.5551.16 ± 5.3939.60 ± 4.7033.05 ± 2.41
564.93 ± 3.8151.31 ± 2.4851.51 ± 5.4739.95 ± 4.7332.87 ± 2.40
665.04 ± 3.8051.37 ± 2.4751.73 ± 5.5340.07 ± 4.7132.81 ± 2.42

Table R3: Performance across varying relation strengths η\eta on CoraFull (%)

η\etaMean AccFinal AccMean NovelFinal NovelPD
063.93 ± 0.7153.71 ± 1.1952.47 ± 1.7047.70 ± 2.1726.59 ± 3.58
165.07 ± 0.9955.58 ± 1.0256.31 ± 1.5451.36 ± 2.3024.63 ± 3.54
265.28 ± 0.9955.86 ± 1.0556.47 ± 1.8651.65 ± 2.1924.35 ± 3.56
365.46 ± 0.9556.04 ± 1.1056.58 ± 2.2951.83 ± 2.2324.17 ± 3.68
465.50 ± 0.9356.18 ± 1.1256.48 ± 2.6651.93 ± 2.3624.02 ± 3.71
565.52 ± 0.8756.37 ± 1.1656.37 ± 2.8652.11 ± 2.3523.83 ± 3.75
665.53 ± 0.8656.39 ± 1.0756.24 ± 3.0152.08 ± 2.2423.82 ± 3.68
  1. Decisive Empirical Gains over η=0\eta = 0: Compared to η=0\eta = 0, enabling MRF prior injection brings marked improvements across all metrics.
  2. Growth Trend: On CoraFull, accuracy jumps sharply at η=1\eta = 1 and largely saturates beyond η≥2\eta \ge 2; on ogbn-arxiv, a similarly strong rise occurs for η≤2\eta \le 2 before plateauing.
  3. Insensitivity to Bin Count BB vs. Naive Homophily (B=1B = 1): As shown in Fig. 6d, the model is robust to the number of bins BB. However, even with only a single bin, it is fundamentally different from a standard homophily prior because it is empirically computed from the base session rather than heuristically assuming positive homophily.

Thank you for the detailed response and additional experiments. The explanation of the discrepancy with Mecoin based on different class splits is reasonable, and the additional η-sensitivity experiments provide useful evidence for the effectiveness of MRF-based refinement.

However, the contribution of the proposed similarity-dependent topological prior is still not fully established, as the η=0 variant retains unary pseudo-labeling and analytic reinjection. Moreover, the claimed ability to handle heterophily remains empirically unverified, since all datasets are homophilous. Overall, while the response addresses some of my concerns, the remaining issues prevent me from changing my original rating.

Reviewer_2HxmOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes MAGIC, a replay-free framework for Graph Few-Shot Class-Incremental Learning (GFSCIL). It combines a frozen graph backbone with analytic continual learning, introduces a topology-aware prior with Potts MRF inference to refine pseudo-labels for novel classes, and employs drift-aware distillation to mitigate representation changes caused by cross-session edges. Experiments on five graph datasets demonstrate promising improvements in classification accuracy and training efficiency over existing baselines.

Critical strengths
  1. The paper addresses graph few-shot class-incremental learning, where learning new classes from limited supervision while retaining historical knowledge remains challenging.

  2. The combination of topology-aware prior learning, MRF-based pseudo-label refinement, and drift-aware analytic distillation provides a coherent, replay-free framework. The use of closed-form updates also offers computational efficiency.

  3. The paper demonstrates strong performance across five datasets under 1-, 3-, and 5-shot settings. The evaluation includes multiple backbones, ablation studies, and efficiency comparisons, providing useful evidence for the effectiveness of the overall framework.

Critical weaknesses

1. Lack of a direct forgetting metric.

Since this work addresses graph few-shot class-incremental learning, evaluating catastrophic forgetting is essential. However, the paper mainly reports Mean Accuracy (MA), Final Accuracy (FA), and Performance Drop (PD), without explicitly reporting a standard forgetting metric.

PD measures the overall accuracy drop between the base and final sessions, which is not equivalent to forgetting. In particular, PD conflates the degradation of previously learned classes with the difficulty of recognizing newly introduced classes. A model may exhibit a large PD even when its old-class accuracy remains unchanged. Therefore, PD cannot serve as a substitute for a task-wise forgetting metric.

Could the authors report Average Forgetting or Backward Transfer (BWT), along with old-class accuracy across incremental sessions? This is particularly important for validating the proposed drift-aware distillation mechanism, whose primary motivation is to preserve historical knowledge.

2. The stated technical challenges are insufficiently specific to establish the paper's novelty.

The first challenge, novel-class overfitting due to limited labeled samples, is a fundamental characteristic of few-shot class-incremental learning rather than a specific unresolved challenge identified by this work. Existing GFSCIL methods, including GFCIL, Geometer, Mecoin, and TAP, are already designed to address learning from extremely limited novel-class supervision.

While the challenge itself is valid, its current formulation does not clearly identify the specific limitations of existing approaches. I think a stronger motivation would explain which aspects of novel-class overfitting remain inadequately addressed by prior methods, why existing solutions are insufficient, and what specific gap motivates the proposed topology-aware prior.

Overall, the broadly stated challenges do not sufficiently convey the novelty of the proposed method. I encourage the authors to formulate more specific, technically grounded research gaps and clarify how the proposed contributions address limitations that existing methods have not adequately resolved.

3. Limited backbone evaluation and unclear generalizability to standard GNN architectures.

Widely used trainable GNN architectures such as GCN, GAT, and GraphSAGE are not included in the backbone. Although the proposed framework is designed around frozen representations, these GNNs could also be trained during the base session and subsequently frozen during incremental learning.

Why are standard GNN architectures excluded from the backbone evaluation? Can the proposed method be effectively applied to frozen GCN, GAT, or GraphSAGE encoders?

4. Could the authors justify why the Potts MRF formulation is appropriate under evolving class and graph structures? Under what conditions can the base-session relational prior reliably transfer to novel sessions? It would be valuable to investigate whether the method remains effective when the homophily/heterophily patterns or class relationships change across sessions. The paper should explicitly evaluate whether the learned relational prior remains valid for cross-session edges, rather than assuming that a relation model estimated from base-session edges transfers reliably to newly introduced classes.

Questions for authors

See weakness.

6 comments
AuthorsOpenReview ↗

Weakness 1

Weakness 1

AF, BWT, and CL Matrix

  • Mean Accuracy (MA), Final Accuracy (FA), and Performance Drop (PD) are the standard evaluation metrics across existing GFSCIL literature (e.g., GFCIL, Geometer, Mecoin, TAP). We adhered to this protocol to ensure a consistent and fair comparison with prior baselines.
  • Here, we provide the Continual Learning performance evaluation matrix on ogbn-arxiv to directly demonstrate forgetting dynamics:
S0S1S2S3S4S5S6S7S8S9S10
After S083.30
After S180.9383.13
After S277.8782.9734.10
After S376.9958.3036.1847.27
After S476.0657.4334.9343.0929.74
After S575.7356.9633.0642.8929.1035.76
After S675.4755.4233.4742.1629.5835.7664.55
After S774.8254.6233.3742.5529.4236.8764.4629.71
After S874.4853.6933.5842.4529.1837.5362.7728.5523.05
After S974.1852.5632.0143.2328.7940.1864.2127.9322.9926.72
After S1073.8151.8231.7042.4528.9539.9662.3427.4620.1326.644.91
  • In the final version, we will include these data across all datasets and baselines:
    1. Average Forgetting (AF) and Backward Transfer (BWT).
    2. Continual Learning (CL) performance evaluation matrix.
AuthorsOpenReview ↗

Weakness 2

Weakness 2

Specificity of Technical Challenges and Novelty

  • Clarification of Research Gap: The research gap we emphasize is not the presence of novel-class scarcity itself (Challenge 1), but the limitations of existing methodologies designed to resolve it.
  • Limitation of Existing Approaches: Existing GFSCIL methods (such as Geometer, Mecoin, and TAP) tackle data scarcity primarily within the node-feature space, and fail to exploit the topological properties of graphs to mitigate novel-session overfitting.
  • Our Novelty: MAGIC estimates a topological prior during the base session and injects it into the model via Potts Markov Random Field (MRF) inference during novel sessions, thereby fully exploiting graph topological properties to alleviate novel-session overfitting.
AuthorsOpenReview ↗

Weakness 3

Weakness 3

Backbone Selection and Trainable GNN Architectures

MAGIC can be directly applied to standard parameterized GNN encoders trained on the base session and subsequently frozen during incremental sessions. We evaluate MAGIC with base-trained and frozen GCN and GAT backbones on CoraFull (5-shot):

Method / BackboneMA (%) ↑\uparrowFA (%) ↑\uparrowPD (%) ↓\downarrow
TAP59.27 ±\pm 1.3545.06 ±\pm 1.9638.48 ±\pm 4.04
MAGIC (GCN)58.80 ±\pm 0.5748.91 ±\pm 1.5229.38 ±\pm 4.97
MAGIC (GAT)57.90 ±\pm 1.0647.80 ±\pm 1.1430.41 ±\pm 4.43
MAGIC (SGC)65.45 ±\pm 0.9556.02 ±\pm 1.0524.19 ±\pm 3.65

As shown in the table, using base-trained and frozen GCN or GAT as the backbone still outperforms the strongest baseline (TAP), demonstrating MAGIC's compatibility with standard GNN architectures, yet underperforms the parameter-free SGC backbone. This gap arises from two main reasons.

  1. GCN and GAT are trained solely on the base session and then frozen, the learned feature representations may not generalize well to novel-class nodes, since our setting has relatively few base classes, which makes the scenario more challenging.
  2. Training GCN and GAT on base session distorts the feature space, so cosine similarities among novel-class nodes lose semantic consistency with base classes, invalidating the transferred topological prior.

By contrast, parameter-free propagation operators and general-purpose pretrained graph foundation models avoid base-class specialization and preserve consistent similarity geometry across base and novel classes. We will include these results and analysis in the revised paper.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary Response

We sincerely thank all reviewers for their constructive feedback. We have addressed each concern through theoretical clarifications, controlled ablations, and additional experiments, summarized below.

To Reviewer 2Hxm — Forgetting Metrics, Novelty Specificity, Backbone Generalizability, and Prior Transferability

We clarified that our evaluation follows the standard GFSCIL protocol and sharpened our research gap: existing methods tackle few-shot scarcity mainly in the node-feature space, whereas MAGIC exploits graph topology via a base-learned Potts MRF prior.

  • Direct forgetting dynamics: We provided the Continual Learning evaluation matrix on ogbn-arxiv showing strong historical retention, and will include Average Forgetting, Backward Transfer, and full matrices across all datasets in the final version.
  • Standard GNN backbones: We experimentally verified that MAGIC also works effectively with trainable GNN backbones such as GCN and GAT trained on the base session and subsequently frozen, while explaining why parameter-free and pretrained backbones better preserve similarity geometry.
  • Prior transferability: We clarified that positive transfer depends on the consistency between connected node similarity and label agreement across sessions, and will state this boundary clearly in the revision.

To Reviewer JrFa — Recent Baselines, Split Alignment, Topological Prior Contribution, and Heterophily

We explained the split-configuration difference behind published Mecoin numbers and added new baseline comparisons, ablations, and heterophily evaluations:

  • Recent GFSCIL baselines: We compared MAGIC against additional recent GFSCIL baselines including URPD, G-Scaler, MEGA, and Meta-Tuner on Coauthor CS, demonstrating that MAGIC achieves the best performance across all metrics, and clarified our methodological novelty over these works.
  • Matched-split comparison: Under Mecoin's exact class split on CoraFull, MAGIC surpasses published Mecoin results in both Final Accuracy and Performance Drop.
  • Topological prior isolation: Our component-wise ablation on ogbn-arxiv confirms that the topological prior is the primary driver of performance gains over unary pseudo-labeling and analytic reinjection alone.
  • Heterophilous graph verification: Experiments on the heterophilous benchmark chameleon_filtered show that MRF prior injection improves both overall and novel-class accuracy while learning negative relational coefficients.

To Reviewer B66w — Joint Ablation, Coupling Isolation, Transductive Protocol, and Distillation Analysis

We addressed each concern through targeted ablations and protocol verifications:

  • Complete joint ablation: We provided additional joint ablation experiments on ogbn-arxiv, confirming both the individual and combined effectiveness of the topological prior and distillation over the analytic ridge baseline.
  • Coupling isolation: Our η\eta-sensitivity experiments on ogbn-arxiv and CoraFull confirm steady gains over η=0\eta = 0, while Figure 6d demonstrates robustness at B=1B = 1 and its fundamental difference from heuristic homophily assumptions.
  • Test-set isolation: Strictly excluding evaluation nodes from the candidate set on ogbn-arxiv yields negligible difference, ruling out test-set leakage.
  • Distillation on affected nodes: Evaluating affected historical nodes on ogbn-arxiv confirms that distillation substantially mitigates cross-session representation drift under frozen backbones. We will refine the weighting scheme and discuss remaining limitations in the revised manuscript.

Overall, these additional results reinforce our central contribution: combining frozen graph backbones with closed-form analytic continual learning, drift-aware distillation, and base-learned Potts MRF priors provides an efficient, replay-free solution to novel-session overfitting and cross-session drift in GFSCIL.

We sincerely thank all reviewers and the Area Chair for their time and constructive guidance.

AuthorsOpenReview ↗

Update on Reviewer JrFa's Latest Comment

Regarding the data concern in Reviewer JrFa's latest comment, we have already replied to clarify that there is no dataset misuse in Table R4. The results are indeed evaluated on ogbn-arxiv and are fully self-consistent with Table R2.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.