{"id":"e65e9a96-9b22-4c8b-a3af-9219ca6630a4","arxiv_id":"2412.18316","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A self-supervised graph learning method jointly learns feature and topology augmentations from the data, reducing reliance on hand-crafted augmentations.","lead":"This paper introduces DSGRL, a graph learning method that learns its own data augmentation from graph structure and node features instead of relying on hand-picked heuristics. It reports matching or better performance than several self-supervised baselines on node classification and graph property prediction benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Trained-vs-untrained comparisons do not isolate the learned augmentation; without a frozen-augmenter control, the claim that jointly learned augmentations drive DSGRL's gains is not established.","rationale":"I read the paper in good faith: it proposes a clear, general framework, reports a broad evaluation, includes code, tunes baselines, and is candid about cases where topology augmentation fails on chemical datasets. The central claim, however, is causal: learned augmentations, not heuristics, drive the representation quality. The Random-F and Random-T baselines are useful sanity checks but are not the right control for this causal claim, because they leave the whole model untrained. The reader's weakest_assumption focuses on the topology threshold heuristic preserving semantics; my concern is adjacent but more specific to the evidence for learned augmentations. On several datasets the trained-versus-untrained margin is small, and the paper's own discussion attributes much of the untrained performance to GNN inductive bias, which directly weakens the inference that the learned augmentation is the source of the gains. A frozen-augmenter control would settle this decisively. Because the reader already flagged the untrained-variant issue in the rationale and issued a conditional verdict, my read does not change the verdict: the paper remains a credible conditional acceptance pending sharper ablation evidence.","tokens_in":13903,"tokens_out":6638,"duration_ms":66624,"concrete_test":"Run DSGRL-Feature and DSGRL-Topology with the augmenter parameters (fΘ1/fΘ2 or tΦ) frozen at their random initialization, updating only the shared encoder hΘ and the downstream linear classifier under the same loss in Eq. 6, on all datasets in Tables 3 and 6. Compare against fully trained DSGRL. If the frozen-augmenter control matches trained DSGRL within one standard deviation on most datasets, the learned augmentations are not the causal driver and the central claim should be weakened; if fully trained DSGRL clearly and consistently exceeds the frozen control, the claim that jointly learned augmentations matter is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in Section 2.5 is that DSGRL is \"jointly optimized on both the augmenter and encoder parameters\" and, as a result, \"the learned augmentations are governed by the inherent signal in the data.\" The empirical support for this causal reading is the gap between trained DSGRL and the Random-F/Random-T variants. But those variants are untrained end-to-end, so they simultaneously remove augmentation learning and encoder training; they cannot isolate whether the augmentation parameters are responsible for any gains. In several cases the gap is small: in Table 6, Random-F reaches 77.0 on DD versus 78.0 for Feature, and on IMDB-B, IMDB-M, and COLLAB the trained variants are within roughly one to two points of Random-F or Random-T. In Table 3, Deezer shows Random-T at 56.0 versus Topology at 59.3, and Random-F at 55.4 versus Feature at 58.2. The paper itself acknowledges that untrained variants are strong competitors, consistent with Trivedi et al. [39] that GNN inductive bias compensates for poor augmentations. Without a control that trains the encoder and linear classifier while keeping the augmenter frozen at initialization, the observed gains could come from the added capacity of the feature MLP or topology GNN, from the regularization terms, or from other training dynamics, rather than from learning a semantically suitable augmentation. This is the load-bearing weak point for the paper's headline contribution.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DSGRL, a self-supervised graph representation learning method with learnable feature and topology augmentations. Feature augmentation is implemented by two feed-forward networks that produce augmented feature views, while topology augmentation is implemented by a GNN that computes high-order node similarities, which are thresholded to form a high-order adjacency view. The augmenters are trained jointly with a shared GNN encoder using a VICReg-style invariance, variance, and covariance objective with an additional model-orthogonality regularizer. The method is evaluated on node classification (including heterogeneous graphs) and graph property prediction, comparing against self-supervised baselines and reporting untrained variants Random-F and Random-T. The central claim is that DSGRL automatically learns augmentations from the graph signal, avoiding hand-designed heuristics, while matching or outperforming existing self-supervised baselines.","tokens_in":14209,"tokens_out":3392,"duration_ms":31808,"significance":"If the central claim holds, DSGRL would provide a generally applicable, data-driven alternative to heuristic graph augmentations, with the attraction of being usable for both homogeneous and heterogeneous graphs. The paper is empirically extensive: it uses multiple public benchmarks, follows standard linear-evaluation protocols, tunes baselines, includes untrained controls, and points to source code and an appendix. These are strengths. The main qualification is that the evidence for the claim that learned augmentations drive the gains is incomplete, because the untrained variants are strong competitors on several datasets and no control isolates the effect of augmenter training from encoder training. The conceptual contribution is therefore plausible but needs a sharper experiment before the strongest conclusions are justified.","major_comments":[{"comment":"This is the load-bearing weakness for the paper's headline contribution. Please add the frozen-augmenter ablation or substantially weaken the causal interpretation.","section":"Section 2.5, Tables 3 and 6"},{"comment":"The topology augmenter is only partially learned. The similarity function g(hi,hj)=hi^T hj and the threshold Ek[g(hi,hk)] are fixed modeling choices, and the paper provides no analysis showing that the resulting thresholded high-order network preserves the task-relevant information needed for downstream classification. The paper itself notes that topological augmentation performs poorly on chemical datasets. This does not invalidate the empirical results, but it weakens the claim that the augmentation is 'suitable' and learned from the signal; an ablation varying the similarity function or threshold, or a theoretical justification of the thresholding operation, would make the data-driven claim more convincing.","section":"Section 2.3.2 and Section 5"},{"comment":"The scalability and heterogeneous-graph claims are based on results with no standard deviations or significance tests: Table 4 reports single numbers for Yelp and Reddit, and Table 5 reports point estimates for a single split on IMDB. Given the paper's emphasis on scalability and generality, these claims should be supported with repeated runs or explicit acknowledgment of the limited evidence.","section":"Tables 4 and 5"}],"minor_comments":[{"comment":"The dataset count is inconsistent: Section 3.1 says the NC datasets are 8, while the abstract and conclusion say nine node classification datasets. Please reconcile this count by explicitly counting the heterogeneous IMDB dataset.","section":"Section 3.1 and Conclusion"},{"comment":"There is a typo in the table header: 'Ramdom-T' should be 'Random-T'.","section":"Table 6"},{"comment":"The conclusion states that 'DSGRL surpasses the baseline SOTA approaches,' but in several cases the differences are within one standard deviation or tied (e.g., NCI1 in Table 6 where Feature reports 75.0 vs. GraphCL's 75.2). A more cautious wording such as 'matches or outperforms' would be consistent with the abstract and the displayed results.","section":"Section 5"},{"comment":"The notation A' = t_Phi'(A) = g(H,H) is slightly confusing because g is described as a function on pairs of node vectors, while the matrix A' has entries a'_ij. Please clarify that g is applied elementwise to pairs of rows of H.","section":"Section 2.3.2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript reports a substantial empirical study and the untrained-variant controls are a step in the right direction. The main issue is that the headline causal claim about learned augmentations is not yet supported by the experiments; the requested frozen-augmenter control is feasible within the paper's scope and should be added. There is no concern about fit; the paper is within scope for a graph representation learning venue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know: this is a serious, clearly written empirical paper on graph self-supervised learning, but its central claim—that jointly learning augmentations is what produces the performance—is not actually backed by the experiments. The method itself is new and worth taking seriously.\n\nThe genuinely new pieces are the learned feature augmenter (an FFN producing two feature views) and the learned topology augmenter (a GNN which computes high-order similarities, thresholded against per-node expected similarity). Combining those with a VICReg-style invariance/variance/covariance objective plus an orthogonality regularizer between the two augmenters is also new. The evaluation is broad: nine node-classification datasets (including heterogeneous graphs) and eight graph-property datasets, with standard linear evaluation protocols, reported means and standard deviations, and hyperparameters tuned for all baselines. The authors also include untrained Random-F/Random-T variants and honestly acknowledge that topology augmentation fails on chemical datasets—that kind of transparency is refreshing.\n\nThe soft spot is the one the stress-test note identifies, and it lands. The trained-versus-untrained comparisons (Feature vs Random-F, Topology vs Random-T) simultaneously change both the augmenter parameters and the degree of end-to-end training. A random-initialized augmenter still feeds meaningful signal through the encoder; the gap on several datasets is small (within a point or two), and the paper even cites prior work showing GNN inductive bias can compensate for poor augmentations. There is no frozen-augmenter control, so we cannot tell whether the gains come from learning a good augmentation, from added model capacity, or from the regularization terms. That is the load-bearing weakness for the 'learned augmentation' narrative.\n\nTwo more things. First, the gains over strong baselines are modest and sometimes losses; claiming to 'surpass SOTA' overreaches the data. Second, the paper omits direct comparison with JOAO and Graph Barlow Twins, both automated-augmentation methods that are directly relevant; citing them isn't enough.\n\nWho is this for: anyone working on graph contrastive learning or augmentation. The architecture is plausible and the empirical study is useful even if the causal story is unproven. I would send it to peer review, but with a request for a frozen-augmenter ablation and a toned-down claim. Without the ablation, the headline claim stays speculative.","headline":"Solid empirical graph SSL paper with a genuinely new learned-augmentation design, but the experiments don't actually prove that the learned augmentations are what drive the gains.","tokens_in":14730,"tokens_out":1891,"would_cite":true,"duration_ms":19865,"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 claims that self-supervised graph representation learning can replace hand-crafted augmentations with augmentations learned from the graph's own signal, and reports matching or beating state-of-the-art baselines on node…","keywords":["self-supervised learning","graph representation learning","graph data augmentation","learnable augmentation","contrastive learning","node classification","graph property prediction","heterogeneous graphs"],"falsifier":"Train DSGRL on a labeled dataset, freeze the learned topology augmenter, and replace its weighted edges with a random graph drawn uniformly among node pairs at the same edge density. If random-topology DSGRL matches learned-topology DSGRL on the downstream linear-evaluation task, then the claim that the learned augmentation itself carries the task-relevant signal is falsified; if the learned graph does better, the claim is supported.","tokens_in":13712,"feed_emoji":"🕸️","tokens_out":8795,"duration_ms":73589,"temperature":0.7,"pith_summary":"The paper argues that the main bottleneck in self-supervised graph representation learning is choosing good graph augmentations, and that hand-picked heuristics often destroy task-relevant structure. It proposes DSGRL, a method that learns the augmentation from the graph's own feature and topological signals, jointly with the representation. A feed-forward network learns multi-view feature perturbations, and a GNN-based module learns a high-order similarity graph by keeping edges between nodes whose learned vectors agree strongly. Tested on nine node-classification datasets and eight graph-property datasets, the authors report that the method matches or outperforms state-of-the-art self-supervised baselines and is comparable to semi-supervised methods. If so, self-supervised graph learning would no longer need dataset-specific hand-designed augmentations.","feed_headline":"Self-supervised graphs learn their own augmentations","feed_subtitle":"A data-driven augmenter matches or beats state-of-the-art baselines on 17 graph datasets without labeled data.","key_machinery":"The argument is carried by three components. The feature augmenter is a pair of feed-forward networks $f_{\\Theta_1}$ and $f_{\\Theta_2}$ that map the node feature matrix $X$ into two views $X_1, X_2$. The topology augmenter is a GNN $h_\\Phi$ that computes high-order (multi-hop) node embeddings $H$ from the graph, then applies the threshold rule $a'_{ij} = h_i^\\top h_j$ if $h_i^\\top h_j > \\mathbb{E}_k[h_i^\\top h_k]$, otherwise $0$, producing a weighted high-order adjacency matrix $A'$. A shared GNN encoder $h_\\Theta$ maps the two views to latent representations $Z_1, Z_2$, and the training loss combines an invariance term $\\|Z_1 - Z_2\\|_F$ with regularizers that keep latent dimensions spread out and decorrelated, plus an orthogonality penalty on the augmenter weights so the two feature augmenters do not collapse into one another. Because all these parameters are optimized together, the augmentation is shaped by the graph's own signal rather than chosen from a fixed set of perturbations.","core_discovery":"The central claim is that augmentation for graph self-supervision can be learned, not hand-picked. DSGRL builds two complementary views: feature views produced by two small feed-forward networks applied to the node feature matrix, and a topological view produced by a GNN that embeds nodes and then keeps a weighted edge between two nodes only when their dot-product similarity exceeds the expected similarity to the other nodes. The two views are encoded by a shared GNN, and the whole system is optimized end-to-end with a loss that asks the two representations to agree, keeps their dimensions spread out and mutually independent, and forces the two feature augmenters to be different from each other. The paper reports that this jointly learned augmentation matches or surpasses state-of-the-art self-supervised baselines on homogeneous and heterogeneous node classification and on graph property prediction, and that it remains trainable on large graphs where full-batch contrastive baselines run out of GPU memory.","pith_inferences":["Editorial inference: a dataset-dependent selector that picks between the feature and topology augmenters would probably improve average performance, since the paper reports topology augmentation underperforms on chemical datasets but does not propose such a selector.","Editorial inference: the threshold rule, which compares each node-pair similarity to the expected similarity to all other nodes, is one natural choice among many; testing cosine similarity or a learned metric would show how much the result depends on that specific function.","Editorial inference: because DSGRL requires no meta-paths, the single heterogeneous-graph experiment suggests the approach could carry over to knowledge graphs and other multi-relational networks, but that extension is not yet demonstrated.","Editorial inference: the strength of the untrained variants implies that future augmentation methods should be benchmarked against random augmentations of equal strength; otherwise the encoder's inductive bias can be mistaken for an augmentation gain."],"forward_implications":["One data-driven pipeline can replace dataset-specific augmentation heuristics across homogeneous, heterogeneous, chemical, and social graphs.","The feature and topology augmenters are complementary and can be combined, so a single method can adapt its view generation to the graph type it encounters.","On large graphs such as Yelp and Reddit the method trains where full-batch contrastive baselines exhaust GPU memory, because it avoids negative sampling and large-batch requirements.","The reported node-classification accuracy is close to the semi-supervised references, which suggests label-free pretraining can substitute for much of the labeled training data.","The paper's own untrained random variants are strong competitors, especially on chemical datasets, so the marginal benefit of the learned augmentation should be judged against those random baselines, not only against older heuristics."],"supporting_citations":[{"why":"Supplies the invariance, variance, and covariance regularizers that prevent representation collapse without negative sampling.","marker":"[1]"},{"why":"Provides the public dataset splits and evaluation protocol used for graph property prediction.","marker":"[9]"},{"why":"Provides the heterogeneous graph dataset and the meta-path-based results used as baselines for the IMDB experiment.","marker":"[10]"},{"why":"Uses diffusion-based high-order networks for augmentation, the approach DSGRL's topology augmenter is designed to learn instead of hand-select.","marker":"[14]"},{"why":"Baseline for graph-level representation via mutual information maximization in the graph property prediction comparison.","marker":"[33]"},{"why":"Baseline that learns edge-dropping augmentations adversarially in the graph property prediction comparison.","marker":"[35]"},{"why":"Motivates the method by documenting that heuristic augmentations can destroy task-relevant signal.","marker":"[39]"},{"why":"Baseline corruption-based contrastive method compared on node classification.","marker":"[40]"},{"why":"Baseline that selects augmentations from a fixed set in the graph property prediction comparison.","marker":"[47]"},{"why":"Baseline using adaptive edge removal and feature masking in node classification.","marker":"[53]"}],"fun_headline_variants":["Graphs learn their own augmentations","Self-supervised graphs auto-tune augmentations","No labels, no heuristics: graphs self-augment","Data-driven augmentation for graph learning","Graph self-supervision learns augmentation on the fly"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that edges between nodes whose learned high-order feature dot products exceed the expected similarity form a view that preserves the information needed for the downstream task; the paper provides no proof that this thresholded similarity graph keeps task-relevant semantics, and it notes the topology augmenter performs poorly on chemical datasets.","fun_headline_variants_meta":{"raw":{"variants":["Graphs learn their own augmentations","Self-supervised graphs auto-tune augmentations","No labels, no heuristics: graphs self-augment","Data-driven augmentation for graph learning","Graph self-supervision learns augmentation on the fly"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000288,"raw_usage":{"total_tokens":1709,"prompt_tokens":983,"completion_tokens":726,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":599,"completion_tokens_details":{"reasoning_tokens":656}},"tokens_in":599,"tokens_out":726,"duration_ms":6644,"temperature":1.0,"reasoning_tokens":656,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T04:48:16.439837+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train DSGRL on a labeled dataset, freeze the learned topology augmenter, and replace its weighted edges with a random graph drawn uniformly among node pairs at the same edge density. If random-topology DSGRL matches learned-topology DSGRL on the downstream linear-evaluation task, then the claim that the learned augmentation itself carries the task-relevant signal is falsified; if the learned graph does better, the claim is supported.","supporting_citations":[{"cited_title":"Vicreg: V ariance-invariance- covariance regularization for self-supervised learning, 2021","cited_arxiv_id":null,"evidence_quote":"Supplies the invariance, variance, and covariance regularizers that prevent representation collapse without negative sampling."},{"cited_title":"A fair comparison of graph neural networks for graph classification, 2020","cited_arxiv_id":null,"evidence_quote":"Provides the public dataset splits and evaluation protocol used for graph property prediction."},{"cited_title":"ACM, (apr 2020)","cited_arxiv_id":null,"evidence_quote":"Provides the heterogeneous graph dataset and the meta-path-based results used as baselines for the IMDB experiment."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Baseline for graph-level representation via mutual information maximization in the graph property prediction comparison."},{"cited_title":"Adversarial graph augmentation to improve graph contrastive learning, 2021","cited_arxiv_id":null,"evidence_quote":"Baseline that learns edge-dropping augmentations adversarially in the graph property prediction comparison."},{"cited_title":"Augmentations in graph contrastive learning: Current methodological flaws; towards better practices, 2021","cited_arxiv_id":null,"evidence_quote":"Motivates the method by documenting that heuristic augmentations can destroy task-relevant signal."},{"cited_title":"Hamilton, Pietro Liò, Y oshua Bengio, and R Devon Hjelm","cited_arxiv_id":null,"evidence_quote":"Baseline corruption-based contrastive method compared on node classification."},{"cited_title":"Graph Contrastive Learning Automated","cited_arxiv_id":"2106.07594","evidence_quote":"Baseline that selects augmentations from a fixed set in the graph property prediction comparison."},{"cited_title":"2069–2080, New Y ork, NY , USA, (2021)","cited_arxiv_id":null,"evidence_quote":"Baseline using adaptive edge removal and feature masking in node classification."}],"review_version":1}