{"id":"9fbd810e-772a-4028-b1e0-0f18e8ea26a9","arxiv_id":"2507.17687","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"OGCIL replays old-class embeddings generated by a conditional VAE and rejects unknown nodes via prototype-centered hypersphere boundaries, improving open-set graph class-incremental learning on five benchmarks.","lead":"This paper combines graph class-incremental learning with open-set recognition, so a graph model can keep adding new node classes without forgetting old ones, while flagging nodes from classes it has never seen. It generates fake node embeddings for old and unknown classes and uses a hypersphere loss to keep known classes compact and reject unknowns.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 8 computes the open-set score in the raw GNN embedding space, while Eq. 6 trains prototypes in the CVAE latent space h(z)=mu_phi(z); the mismatch makes the reported OSCR/AUC-ROC depend on an unstated inference-time choice.","rationale":"The reader's weakest assumption is correct and is the most load-bearing issue. Section 3.3.1 defines z as the GNN-generated embedding and h as the VAE latent variable; Eq. (6) explicitly uses h(z) = mu_phi(z) with prototypes p_c, so the classification boundary lives in latent space. Section 3.3.3 generates pseudo ID embeddings by decoding samples from N(p_c, I), placing them in raw z space; Eq. (5) mixes these with current raw embeddings to make pseudo OODs. No operation in the paper maps p_c (or the decision boundary) back to raw z space. Yet Eq. (8), the only scoring rule for test nodes, computes distance from raw z to p_c. This is not a minor typo because OSCR and AUC-ROC are defined over this score; a wrong distance space can arbitrarily reorder nodes and change the reported margins (e.g., up to 17.6% OSCR). The absence of released code means a reader cannot determine whether the implementation actually used h(z) at inference, i.e., whether the written method is complete. We therefore agree with the CONDITIONAL verdict: the paper should either correct Eq. 8 to h(z), state and justify that raw z and h(z) are aligned, or provide code. We do not see a basis to reject outright, since the framework and experiments are otherwise coherent, and the concern is resolvable by a targeted check.","tokens_in":18038,"tokens_out":7150,"duration_ms":72974,"concrete_test":"Implement the method exactly as in Sec. 3 using a public GCN backbone and the CoraFull task splits of Table 2. Train one model, then at test time compute both candidate scores on the same checkpoints: s_raw(z) = -min_c ||z - p_c||^2 (as written in Eq. 8) and s_lat(z) = -min_c ||mu_phi(z) - p_c||^2 (the space consistent with Eq. 6). Recompute OSCR and AUC-ROC for both. If the two score definitions yield materially different OSCR/AUC (e.g., >1 point) or change the ranking against the best baseline, then Eq. 8 is irreconcilable with Eq. 6 and the reported numbers require clarification; if they nearly coincide, the ambiguity is empirically harmless.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The written method is internally inconsistent about which representation feeds the open-set score. In Sec. 3.4.2, Eq. (6) defines the prototypical HSC loss on h(z) = mu_phi(z), the CVAE encoder output, and the prototypes p_c are learned as centers in that latent space. In Sec. 3.3.3, pseudo ID embeddings are produced by decoding samples h ~ N(p_c, I), so they live in the raw GNN-embedding space, and in Sec. 3.4.1, pseudo OOD samples are linear mixes of those raw-space embeddings. Thus all training signals that shape the decision boundary (L_phsc, L_pcvae, and the mixing strategy) operate in the latent h-space after re-encoding for L_phsc, but inference Eq. (8) computes s_open(z) = -min_c ||z - p_c||^2 with the raw GNN embedding z. If raw z and h(z) are not in the same metric space, distances from z to p_c are not meaningful, and the threshold on s_open cannot separate known from unknown nodes. Since the headline gains (OSCR up to +17.6%, higher AUC-ROC) are computed from this score, the central claim rests on an unstated inference-time representation choice. The paper does not release code, so a reader cannot determine whether the experiments actually used h(z) despite the notation in Eq. (8). This is a load-bearing gap, not a stylistic choice.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes OGCIL, a framework for open-set graph class-incremental learning in which a GNN backbone is trained incrementally over tasks with disjoint classes while test nodes may come from classes never seen during training. To mitigate catastrophic forgetting, the method trains a prototypical conditional variational autoencoder whose class-dependent Gaussian prior is centered on learned class prototypes, and generates pseudo in-distribution embeddings by decoding samples from that prior. To detect unknown classes, the method synthesizes pseudo out-of-distribution embeddings by linearly mixing embeddings from different classes, and trains a prototypical hypersphere classification loss that keeps known-class embeddings close to their prototypes while repelling all other samples. The paper reports experiments on five graph benchmarks, comparing against class-incremental and open-set baselines, with OSCR gains up to 17.6% over the best competitor, and includes ablations, hyperparameter studies, and t-SNE visualizations. The central claim is that OGCIL simultaneously reduces catastrophic forgetting and improves open-set detection in this unified setting.","tokens_in":18303,"tokens_out":5261,"duration_ms":54220,"significance":"If the representation-space issue identified below is resolved, this is a useful contribution: it opens a realistic problem setting that combines graph class-incremental learning with open-set recognition, and it provides a systematic empirical study over five datasets with several baselines. The strengths of the paper are its breadth of experiments, the ablations isolating each component (knowledge distillation, prototypical HSC loss, pseudo ID, and pseudo OOD generation), the sensitivity analyses over sample counts, hyperparameters, exemplar strategies, and backbone architectures, and the provision of pseudocode. The proposed method is also modular, as the knowledge-distillation component is stated to be replaceable by other regularization techniques. The main reservation is the load-bearing ambiguity about which representation is used at inference to compute the open-set score, which directly affects the validity of the reported OSCR and AUC-ROC numbers.","major_comments":[{"comment":"The prototypical hypersphere classification loss is defined on the CVAE-encoded representation h(z) = mu_phi(z), so the prototypes p_c are trained in the latent h-space, but the open-set score at inference is written as s_open(z) = -min_c ||z - p_c||^2 using the raw GNN embedding z. Unless raw z and h(z) are shown to live in the same metric space, which the paper does not establish, distances from z to p_c are not meaningful and the threshold on s_open cannot separate known from unknown nodes. Since the headline results in Table 1 are computed from this score, the paper's central claim depends on an unstated inference-time representation choice. Please change Eq. (8) to use h(z) and specify how h is computed at inference, or justify that the two spaces coincide; if the experiments already used h(z), the equation and surrounding text must be corrected accordingly.","section":"Section 3.4.2, Eq. (6), and Section 3.5, Eq. (8)"},{"comment":"Pseudo ID embeddings are generated by sampling h approximately N(p_c, I) and decoding to z_hat = D_theta(z|h), so replay samples live in the raw GNN-embedding space, and the mixing strategy in Eq. (5) operates on those raw-space embeddings. Equation (6) then applies the CVAE encoder h(.) to these pseudo IDs as if they were original embeddings, meaning the replay distribution is sampled in one space while the classification loss is evaluated in another. This double representation shift is neither stated nor justified; please clarify whether pseudo IDs are re-encoded before entering L_phsc and report the effect of this choice on the ablations shown in Figure 3.","section":"Section 3.3.3 and Section 3.4.1"},{"comment":"The appendix states that the code will be released on GitHub upon publication, and the manuscript does not specify which representation was actually fed to the open-set score in the experiments beyond the formula in Eq. (8). Given the mismatch identified above, the reported OSCR and AUC-ROC improvements cannot be independently verified from the text. Please provide the exact evaluation procedure, including the representation used for the open-set score, the thresholding details, and ideally the code or a precise pseudo-code line for the inference step.","section":"Section 6.1 and overall reproducibility"}],"minor_comments":[{"comment":"The reconstruction term is weighted by lambda_reconst but the KL term is unweighted; please clarify whether the KL term is intentionally unweighted or should also carry a balancing coefficient.","section":"Section 3.3.1, Eq. (1)"},{"comment":"Line 8 of the pseudocode, 'Generate H ID embeddings via D_theta(.), {p_c}', contains a typo and should read 'Generate N_ID pseudo ID embeddings' using the variable N_ID declared in the input list.","section":"Section 6.1, Algorithm 1"},{"comment":"For CoraFull, the per-task known and unknown counts sum to more than the dataset's 45 classes across five tasks; please clarify how many classes are permanently unknown versus unknown in earlier tasks and known later, so that the splits are unambiguous.","section":"Table 2"},{"comment":"In the paragraph beginning 'In comparison with other open-set baselines', the phrase 'Our HSC loss' should be lowercase ('our HSC loss') because it appears mid-sentence.","section":"Section 4.2"},{"comment":"The sentence 'Whereas excessively large numbers introduce noise' is a fragment; please join it to the preceding sentence for readability.","section":"Section 4.4.1"},{"comment":"The claim of being the first to unify GCIL and OSR should be sharpened in light of the discussed OpenWRF and LifeLongGNN methods, by stating explicitly how the proposed setting differs from those works (e.g., no full historical data access).","section":"Section 1 and Section 2.2"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a timely and underexplored combination of class-incremental learning and open-set recognition on graphs, and the experimental study is extensive. The central correctness issue is the mismatch between the h-space in which the prototypical HSC loss and prototypes live and the raw z-space used in the written open-set score; this is load-bearing for the reported OSCR and AUC-ROC numbers. I recommend requesting a clear statement of the actual inference-time representation, a corrected Eq. (8), and ideally code release before acceptance. If the experiments used h(z), the paper is likely acceptable after a careful revision; if they used raw z as written, the experimental results would need to be re-examined. The 'first attempt' novelty claim should also be tempered or carefully scoped."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is worth reading for the problem alone: it defines open-set graph class-incremental learning, a combination I have not seen before, and proposes a sensible first cut at solving it. The idea of replaying pseudo-ID embeddings via a prototypical CVAE and synthesizing pseudo-OOD samples by mixing is genuinely new in the graph domain, and the prototypical hypersphere loss is a reasonable way to keep unknowns outside the known-class boundaries. The experiments cover five datasets and a wide set of baselines, and the ablations show each component matters. The authors are honest that their closed-set accuracy is not always best, which makes the OSCR gains more credible.\n\nThe trouble is the representation-space mismatch that you flagged. In Eq. (6), the HSC loss operates on h(z) = mu_phi(z), the CVAE encoder output, and the prototypes p_c are trained as centers in that latent space. But Eq. (8) computes the open-set score as -min_c ||z - p_c||^2 with the raw GNN embedding z. Unless z and h(z) live in the same metric space, which the paper never claims or justifies, the distances in Eq. (8) have no clear meaning. This is not a stylistic quibble; it is the exact score used to separate known from unknown nodes and to produce the reported OSCR and AUC-ROC numbers. The paper does not release code, so I cannot check whether the experiments actually used h(z) in Eq. (8) despite the notation. The stress-test note is right on target.\n\nThe rest of the method is coherent, and the problem formulation is solid. I found no circularity or serious citation issues. The missing error bars and significance tests are also a concern, but secondary to the Eq. (6)/Eq. (8) gap. If the authors intended h(z) in Eq. (8), a one-line fix might resolve everything; if they really used raw z, the method as written is invalid for open-set detection as described.\n\nMy bottom line: this deserves a serious referee, because the problem is timely and the framework is promising. But I would not cite it in its current form, and the paper needs major revision or a clear release of code to resolve the mismatch before the experimental claims can be believed. I would bring it to my reading group mainly to discuss the ambiguity and what it implies for prototype-based OSR in incremental settings.","headline":"New problem framing, but a representation-space mismatch in the written method makes the headline OSR numbers hard to trust without clarification.","tokens_in":18928,"tokens_out":1921,"would_cite":false,"duration_ms":22628,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proposes OGCIL, unifying graph class-incremental learning and open-set recognition: a prototypical CVAE replays old-class embeddings while a hypersphere loss rejects unknowns, yielding up to 17.6% higher open-set classification…","keywords":["graph neural networks","class-incremental learning","open-set recognition","out-of-distribution detection","catastrophic forgetting","variational autoencoder","prototype learning","node classification"],"falsifier":"Remove the space mismatch by evaluating OGCIL with the open-set score computed on the encoded embedding $h(z)=\\mu_\\phi(z)$ instead of the raw embedding $z$ in Eq. 8 on CoraFull; if open-set AUC-ROC changes materially, the published score is measuring distance in the wrong space and the method's rejection regions are not where the paper says they are. A simpler check: measure the mean distance $\\|z-h(z)\\|$ for known-class test nodes relative to the learned hypersphere radii; if it is comparable to or larger than the radii, Eq. 8 cannot be thresholding the same geometry the loss trained.","tokens_in":17742,"feed_emoji":"🕸️","tokens_out":9058,"duration_ms":83995,"temperature":0.7,"pith_summary":"Graph class-incremental learning has assumed closed worlds where every test node belongs to a known class. This paper argues that real deployments also face unknown classes, and that forgetting and open-set failure reinforce each other: forgotten old classes get mislabeled as unknown, while poorly handled unknowns destabilize what the model remembers. It introduces OGCIL, which tackles both at once by generating pseudo-embeddings: a prototypical conditional variational autoencoder replays old-class embeddings, a mixing strategy synthesizes out-of-distribution samples, and a prototypical hypersphere classification loss keeps each known class inside its own prototype-centered ball while pushing everything else outside. Across five graph benchmarks, OGCIL reports open-set classification rates up to 17.6 percentage points above the strongest competitor, with higher open-set AUC-ROC, suggesting the two goals can indeed be met in one framework.","feed_headline":"One framework stops graph forgetting and detects unknown classes","feed_subtitle":"Pseudo-embedding replay plus prototype-centered rejection tops open-set scores on five graph benchmarks.","key_machinery":"The load-bearing object is the prototypical hypersphere classification loss $L_{\\mathrm{phsc}}$, built on the radial-basis similarity $\\varphi(h(z),p_c)=\\exp(-\\|h(z)-p_c\\|^2)$ between the CVAE-encoded embedding $h(z)=\\mu_\\phi(z)$ and each learnable class prototype $p_c$. For a sample of class $c$, the loss maximizes similarity to $p_c$ and minimizes it for every other prototype, so mixed pseudo-OOD samples and off-class samples are pushed outside all hyperspheres. The prototypical CVAE supports this by imposing a Gaussian latent prior $\\mathcal{N}(p_c,I)$ per class, so decoded pseudo-IDs inherit the same centers, and the mixing step creates OOD samples that lie off all prototypes. Knowledge distillation over real and exemplar embeddings stabilizes the GNN representation so generated pseudo-samples remain valid across tasks.","core_discovery":"The central claim is that a single framework can simultaneously mitigate catastrophic forgetting and detect unknown classes in graph class-incremental learning, a setting the paper says no previous method addresses. OGCIL decouples node representations into a task-stable part, regularized by knowledge distillation, and a task-variant encoder. A prototypical CVAE with latent prior centered on class prototypes generates pseudo in-distribution embeddings for old classes, enabling replay without raw graph storage; mixing these with current embeddings produces pseudo out-of-distribution samples. The prototypical hypersphere classification loss then anchors encoded in-distribution embeddings to their class prototypes via a radial-basis similarity and repels off-class and OOD samples, explicitly modeling unknowns as outliers instead of a single unknown cluster. The paper reports consistent OSCR and AUC-ROC gains over class-incremental and open-set baselines across CoraFull, Computer, Photo, CS, and Arxiv, with ablations indicating each component is necessary.","pith_inferences":["A testable extension is to compute the open-set score on the encoded embedding $h(z)=\\mu_\\phi(z)$ instead of the raw embedding $z$; if results change materially, the paper's stated score is measuring distance in a different space than the one the loss trains.","The paper leaves implicit that the same embedding-level generation could transfer to non-graph continual learning or to task-incremental variants, since it never depends on reconstructing the raw input structure.","The dataset protocol, where unknowns of one task become known later, invites a follow-up evaluation of whether the model can re-identify previously rejected nodes as known, testing the interplay of rejection and memory.","Mixing-based OOD generation with a Beta coefficient could be extended to pair-class or adversarial mixup to better cover boundary regions and further sharpen rejection regions."],"forward_implications":["Old-class knowledge can be replayed without storing raw graphs, so continual learners become lighter and more privacy-friendly.","Unknown samples are rejected per prototype rather than forced into one 'unknown' class, which should generalize better to multiple unseen classes.","The framework stays effective with as few as one exemplar per class and with different GNN backbones, suggesting it transfers across memory and architecture constraints.","Thresholding the open-set score $-\\min_c \\|z-p_c\\|^2$ gives a simple deployment rule for flagging novel nodes at inference time.","Because replay happens in embedding space, the approach can be layered on top of existing exemplar-selection and regularization strategies."],"supporting_citations":[{"why":"Supplies the conditional variational autoencoder formulation whose class-dependent latent prior the paper re-centers on learnable prototypes.","marker":"[35]"},{"why":"Motivate the prototype-centered hypersphere loss and the view of unknowns as outliers rather than a single cluster.","marker":"[31, 32]"},{"why":"Provides the coverage-maximization exemplar selection and the ER-GNN replay baseline that OGCIL builds on and is compared with.","marker":"[50]"},{"why":"Introduces proxy-unknown generation via mixup (G2Pxy), the approach the paper's mixing-based OOD generation extends and contrasts with.","marker":"[48]"},{"why":"Open-world graph learning baseline that maximizes uncertainty for low-confidence samples; its inductive-setting failure motivates the new framework.","marker":"[45]"},{"why":"Recent graph class-incremental method TPP, a strong GCIL competitor the paper compares against.","marker":"[25]"},{"why":"Defines OSCR, the threshold-agnostic open-set classification metric used to report the main empirical gains.","marker":"[5]"},{"why":"Classic EWC regularization baseline used in the comparison to show forgetting-prevention approaches do not handle unknowns.","marker":"[15]"}],"fun_headline_variants":["Open-set graph learning: forget less, spot unknowns","Graph GNNs that learn new classes without forgetting old","Prototype replay plus OOD rejection fixes graph forgetting","One graph model tackles new classes and unknown detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's open-set score compares a raw node embedding to prototypes, while the training loss compares an encoded embedding to those same prototypes, so the whole detection scheme assumes these two spaces are effectively the same.","fun_headline_variants_meta":{"raw":{"variants":["Open-set graph learning: forget less, spot unknowns","Graph GNNs that learn new classes without forgetting old","Prototype replay plus OOD rejection fixes graph forgetting","One graph model tackles new classes and unknown detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00023,"raw_usage":{"total_tokens":1526,"prompt_tokens":1031,"completion_tokens":495,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":647,"completion_tokens_details":{"reasoning_tokens":432}},"tokens_in":647,"tokens_out":495,"duration_ms":5843,"temperature":1.0,"reasoning_tokens":432,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:20:19.283326+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Remove the space mismatch by evaluating OGCIL with the open-set score computed on the encoded embedding $h(z)=\\mu_\\phi(z)$ instead of the raw embedding $z$ in Eq. 8 on CoraFull; if open-set AUC-ROC changes materially, the published score is measuring distance in the wrong space and the method's rejection regions are not where the paper says they are. A simpler check: measure the mean distance $\\|z-h(z)\\|$ for known-class test nodes relative to the learned hypersphere radii; if it is comparable to or larger than the radii, Eq. 8 cannot be thresholding the same geometry the loss trained.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the conditional variational autoencoder formulation whose class-dependent latent prior the paper re-centers on learnable prototypes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the coverage-maximization exemplar selection and the ER-GNN replay baseline that OGCIL builds on and is compared with."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces proxy-unknown generation via mixup (G2Pxy), the approach the paper's mixing-based OOD generation extends and contrasts with."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Open-world graph learning baseline that maximizes uncertainty for low-confidence samples; its inductive-setting failure motivates the new framework."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Recent graph class-incremental method TPP, a strong GCIL competitor the paper compares against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines OSCR, the threshold-agnostic open-set classification metric used to report the main empirical gains."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Classic EWC regularization baseline used in the comparison to show forgetting-prevention approaches do not handle unknowns."}],"review_version":1}