{"id":"d2f30e52-bf14-457d-8851-724db855fda6","arxiv_id":"2412.14451","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"CLDG applies contrastive learning to dynamic graphs by sampling multiple timespan views and pulling together the same node's representations across views, outperforming 12 baselines on 7 datasets.","lead":"This paper introduces CLDG, an unsupervised method that learns dynamic graph representations by contrasting the same node across different time-window views, using an inductive bias called temporal translation invariance. It reports state-of-the-art accuracy on seven datasets while using far fewer parameters and less training time than existing dynamic graph models.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table III's SOTA claim is not yet verifiable: the inference protocol is unspecified, temporal leakage is not ruled out, and three metrics contradict the 'outperforms unsupervised baselines' statement.","rationale":"The reader's conditional verdict is reasonable, but the most load-bearing issue is not the philosophical validity of temporal translation invariance. The method can work even if that invariance is only approximate; what makes the headline claim untestable is the missing inference protocol and the possibility that future information enters the test embeddings. The paper does provide code, which is real evidence in its favor, and the method is simple and plausible; no claim of dishonesty is intended. The concrete discrepancy in Table III (three metrics where an unsupervised baseline beats CLDG) is objective and should be corrected regardless of the underlying method's merit. A conditional acceptance is appropriate: the authors should specify the inference-time view construction, clarify the train/validation/test split semantics, report the temporally clean results, and correct the overstatement about outperforming all unsupervised baselines. These fixes would settle whether the central claim survives.","tokens_in":18714,"tokens_out":10102,"duration_ms":90659,"concrete_test":"Using the released code, run CLDG on Bitcoinotc under three explicitly defined inference protocols: (1) embeddings from a single random timespan view, as in training; (2) embeddings from the full temporal graph; (3) embeddings from only the first 80% of the timeline, with test nodes' future edges removed. Report Accuracy and Weighted-F1 for each. If protocol (3) drops materially below the Table III entry, or if protocols (1) and (2) differ by more than a few points, the claimed SOTA is an artifact of the unspecified evaluation protocol. Independently recount the strictly-best unsupervised metrics in Table III and compare with the paper's claim that both CLDG variants outperform all unsupervised baselines.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central empirical claim rests entirely on Table III, but the evaluation protocol that produced it is under-specified in a way that can change the result. Section V-A3 says only that datasets are split 1:1:8 and that after unsupervised training the authors 'freeze the model parameters and output the learned representations for all nodes.' It never states whether those embeddings come from the full temporal graph, from one sampled timespan view, or from a fixed aggregation of views, and it never states whether the 1:1:8 split is by node or by time. Since CLDG training samples timespan views from the whole timeline (Section IV-B, Eqs. 1-5), the embedding used for linear probing can in principle incorporate edges and nodes from the test interval. If the split is by time, that is direct future leakage; if the split is by node, each test node's full temporal neighborhood is still visible during unsupervised pretraining. The table also does not match the text: CAW beats CLDGnode on Bitcoinotc Accuracy (59.85 vs 59.17), DGI beats CLDGnode on BITalpha Weighted-F1 (73.19 vs 72.90), and MNCI beats CLDGnode on Reddit Weighted-F1 (65.26 vs 62.87). Therefore the statement that 'both of which outperform other unsupervised methods' is false as written, and the reported SOTA is not reproducible from the paper without additional protocol details.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CLDG, an unsupervised contrastive learning framework for dynamic graphs. The method assumes that node semantics and labels are approximately stable across different time intervals ('temporal translation invariance'), samples multiple timespan views of a dynamic graph, and trains a shared encoder with InfoNCE-style local and global contrastive losses. The authors report experiments on seven dynamic graph datasets and twelve baselines, claiming state-of-the-art unsupervised performance and competitiveness with supervised methods, while reducing model parameters and training time by large factors relative to existing dynamic-graph models.","tokens_in":19039,"tokens_out":3668,"duration_ms":30355,"significance":"If the empirical claims hold, CLDG would be a simple, lightweight, encoder-agnostic contrastive approach that unifies discrete-time and continuous-time dynamic graph representation learning, with strong practical appeal because it avoids sequence models and reconstructing future states. The paper includes extensive experiments, an ablation of sampling strategies, and a public code/data repository, which are valuable. However, the current evidence does not yet establish the central SOTA claim: the evaluation protocol is under-specified, some reported numbers contradict the text, and the core temporal translation invariance assumption is asserted rather than quantitatively validated.","major_comments":[{"comment":"The evaluation protocol that produces Table III is under-specified in a way that could change the results. The paper states a 1:1:8 train/validation/test split but does not say whether the split is by node or by time, and it says only that after unsupervised training the authors 'freeze the model parameters and output the learned representations for all nodes.' It is not stated whether those embeddings come from the full temporal graph, from one sampled timespan view, or from an aggregation of views. Since CLDG training samples timespan views from the whole timeline (Section IV-B, Eqs. 1-5), a time-based split would make the test interval visible during pretraining, while a node-based split still exposes each test node's complete temporal neighborhood. Please specify the split direction, the exact embedding source used for linear probing, and confirm that no test-interval edges or nodes enter the unsupervised training phase.","section":"V-A3"},{"comment":"The sentence 'both of which outperform other unsupervised methods' is false as written. In Table III, CAW achieves 59.85 Accuracy on Bitcoinotc versus CLDGnode's 59.17; DGI achieves 73.19 Weighted-F1 on BITalpha versus CLDGnode's 72.90; and MNCI achieves 65.26 Weighted-F1 on Reddit versus CLDGnode's 62.87. The later statement that eleven metrics are optimal among the unsupervised methods is consistent with the table, but the blanket claim that both CLDG variants outperform all unsupervised baselines is not. Please correct the claim and report the exact set of metrics in which CLDG is and is not optimal.","section":"V-B, Table III"},{"comment":"The entire method rests on the temporal translation invariance assumption, but it is not quantitatively established. Section IV-A reports only a qualitative observation (Figure 1) and states that prediction labels of the same node tend to be similar in different timespans 'regardless of the encoder used,' without giving any numerical agreement measure across the seven datasets. Because this assumption directly justifies the positive pairs in Eqs. 9-12, and the conclusion admits the method may fail when graph changes are non-continuous or labels constantly change, please provide a quantitative study of label/representation stability across timespan pairs, per dataset, and characterize the regime in which the assumption breaks down.","section":"IV-A"},{"comment":"Table III does not report the timespan view sampling configuration used for the main results. The paper introduces four sampling strategies, the view timespan factor s, and the number of views v, and Section V-C shows that these choices can change accuracy by about 1.4% on DBLP and 4.0% on TAX. Without stating which strategy and which s and v values were used for each dataset in Table III, the reported gains are not reproducible and could be configuration-dependent. Please report the exact configuration (or a pre-specified selection rule) used for the headline results.","section":"V-A4 and IV-B"}],"minor_comments":[{"comment":"The abstract reports an average parameter reduction of 2,001.86 times, while the conclusion reports 'up to 2,000.86 times'; Table V shows per-dataset ratios that vary substantially. Please clarify whether the quoted number is an average or a maximum and define the averaging set.","section":"VI"},{"comment":"The text refers to 'Table Va' and 'Table Vb' but the captions are labeled 'Table IV'; the table numbering should be made consistent.","section":"V-C"},{"comment":"Equation (5) uses set notation for the interval containing |T_i - T_{i±1}| but does not specify how the tuple ordering is enforced in the random sampling strategy; please clarify the sampling procedure.","section":"IV-B"},{"comment":"For the high and low overlap strategies, the stated ranges for T1 (e.g., [min(T) + dt/(2s), max(T) - (2+v)·dt/(4s)]) implicitly assume a particular ordering and may be invalid for large v; the constraints should be stated precisely over the whole tuple.","section":"IV-B"},{"comment":"The statement that CLDGnode 'outperforms the previous state-of-the-art GraphSAGE model by 1.47% on the average of all metrics' should specify whether this is averaged over datasets and both Accuracy and Weighted-F1, and should report the standard deviation or significance testing, as differences of this size may be within run-to-run variation.","section":"V-B"}],"recommendation":"major_revision","confidential_remarks":"The central concern is not circularity but verifiability: the main experimental claim is not reproducible from the paper because the evaluation protocol is incomplete, and the text overstates the table's results. The method itself is plausible and lightweight, and the core idea is worth pursuing, but the authors need to add protocol details, correct the contradictory statements, and provide quantitative evidence for the temporal translation invariance assumption before the SOTA claim can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: clean, simple contrastive method for dynamic graphs, with code and data out, but the headline SOTA claim is not yet verifiable. I checked Table III, and the stress-test counts are right: the table contradicts the text on several cells, and the evaluation protocol is under-specified enough that temporal leakage is not ruled out.\n\nWhat is actually new: the timespan view sampling layer with four overlap-control strategies, and the local/global temporal translation invariance objectives (same-node pairs, node-neighborhood pairs across time windows). This is not a deep theoretical departure—InfoNCE over time-window views of the same graph is a natural fit—but it is a real, dynamic-graph-specific way to pick contrastive pairs, and the engineering is sensible. The parameter/time comparisons in Table V are striking, and the encoder-agnostic check with GCN, GAT, and GraphSAGE is the right kind of robustness evidence. Shipping code and data is also real evidence.\n\nThe soft spots are mostly around the empirical foundation. Section IV-A asserts temporal translation invariance based on empirical studies but never reports the numbers; given that every positive pair in the method depends on this property, it deserves at least a quantitative table. More importantly, Section V-A3 does not say whether the 1:1:8 split is by node or by time, nor does it say which embeddings are frozen and fed to the linear classifier. Because CLDG samples timespan views from the whole timeline during training, the final test interval can in principle leak into the learned representations. That makes the SOTA claim unverifiable as written.\n\nThere is also a direct inconsistency in the text. The claim that both CLDG variants outperform other unsupervised methods is false in Table III: CAW wins Bitcoinotc Accuracy, DGI wins BITalpha Weighted-F1, and MNCI wins Reddit Weighted-F1 over CLDGnode. Counting all fourteen dataset-metric cells, the two CLDG variants take eleven, which is strong but not a clean sweep.\n\nThe conclusion is honest: it concedes the method fails when graph changes are non-continuous or labels constantly change. That is a real scope restriction and should be kept in the final version.\n\nWho this is for: people who want a cheap, simple unsupervised baseline for dynamic graphs. It is not field-redefining. With code available and broad experiments, I would not desk reject. It deserves a serious referee, but the revision must specify the split protocol, report the exact sampling configuration used for Table III, and correct the overstatement.","headline":"Clean, simple dynamic-graph contrastive method with real engineering value, but the SOTA claim is not yet verifiable because the evaluation protocol is under-specified and Table III contradicts its own text.","tokens_in":19488,"tokens_out":4995,"would_cite":false,"duration_ms":40072,"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":"The paper argues that the temporal dimension of a dynamic graph can supply contrastive learning with positive pairs, replacing perturbation-based augmentation.","keywords":["dynamic graph","contrastive learning","temporal translation invariance","graph representation learning","unsupervised learning","timespan view sampling","graph neural networks","node classification"],"falsifier":"Take a dataset where node labels demonstrably change over time, such as a citation network re-labeled as researchers change fields or a synthetic temporal graph with periodic label flips. If CLDG's accuracy degrades toward or below a static contrastive baseline as the label-change rate rises, the temporal translation invariance assumption is the cause. A direct check is to compute same-node label agreement between non-overlapping windows and test whether it predicts CLDG's gain over perturbation-based baselines.","tokens_in":1637,"feed_emoji":"🕰️","tokens_out":2219,"duration_ms":52372,"temperature":0.7,"pith_summary":"The paper claims that dynamic graphs carry a property it calls temporal translation invariance: the semantics and labels of a node tend to stay similar across different time windows, regardless of which encoder is used. CLDG exploits this by sampling several timespan views of the same dynamic graph and treating the same node, or a node and its neighborhood, across those views as positive contrastive pairs instead of corrupting the graph to make views. If this is right, temporal information itself supplies the self-supervised signal, avoiding the semantic damage that perturbation-based augmentations cause. On seven datasets CLDG reports state-of-the-art accuracy among twelve unsupervised baselines and exceeds supervised methods on four datasets, while using about 2,000 times fewer parameters and 130 times less training time than existing dynamic graph models.","feed_headline":"Time-window contrast pairs beat 12 dynamic-graph baselines","feed_subtitle":"A lightweight unsupervised model treats each node's own history as its positive pair, cutting parameters about 2,000x.","key_machinery":"The load-bearing object is the timespan view sampling layer, which converts one dynamic graph into several temporally ordered views by choosing centers on the overall time interval and keeping edges within a window of width $\\Delta t / s$, controlled by a view size factor $s$ and a view count $v$. The invariance assumption it rests on is temporal translation invariance: the same node in different windows is semantically the same entity, so its representations can be pulled together. Four sampling strategies are compared: sequential, high overlap, low overlap, and random. The counterintuitive result is that high overlap, which physically shares the most context, performs worst, while sequential and random sampling perform best. The method then uses InfoNCE to enforce local temporal translation invariance between node embeddings and global temporal translation invariance between a node and its neighborhood embeddings.","core_discovery":"CLDG's central discovery is that the temporal dimension of a dynamic graph is not just a feature to be modeled with recurrent networks but a source of free contrastive views. The authors first demonstrate empirically that nodes' predicted labels are similar across different timespans on their seven datasets, naming this temporal translation invariance. The method then samples $v$ timespan views via a sampling layer, encodes them with a shared-weight GCN, pools neighborhoods with a readout, projects through an MLP, and trains InfoNCE losses that pull the same node's representations across views together while pushing different nodes apart. Both a local variant (node-to-node) and a global variant (node-to-neighborhood) are reported, and the local variant is usually stronger. The reported effect is that an unsupervised model with no labels, no perturbation augmentations, and no sequence model reaches the best unsupervised classification results on all seven datasets and beats the best supervised baseline on four datasets.","pith_inferences":["A direct testable extension is to measure temporal translation invariance on a new dataset before applying CLDG: if label agreement of the same node across non-overlapping windows is low, the positive-pair premise fails and the method should not be expected to help.","The high-overlap result hints that contrastive difficulty, not temporal proximity, is what drives representation quality, which could connect CLDG to curriculum learning and hard-negative selection.","In settings with gradual concept drift, one could anneal the weight of distant timespan views or resample windows per epoch so that positives respect a drift budget while keeping negatives informative.","Because the method only needs node identities to align views, it could be combined with temporal point-process encoders to add future-event prediction without giving up the lightweight contrastive objective."],"forward_implications":["Dynamic graph representation learning can be done without recurrent networks: temporal cues are exploited implicitly by the contrastive objective, so model size and training time scale with the encoder rather than with a sequence model.","Existing static-graph contrastive methods can be adapted to dynamic graphs by replacing perturbation-based view generation with timespan view sampling, avoiding augmentation-induced label changes.","The encoder is swappable: GCN, GAT, and GraphSAGE all give competitive results inside CLDG, so future encoders can be integrated without redesigning the temporal component.","Low-overlap or non-overlapping timespan views are preferable to highly overlapping ones, suggesting that the model improves when the positive-pair task requires semantic rather than physical continuity.","The same sampling layer handles both discrete-time and continuous-time dynamic graphs, giving the method broader coverage than dynamic graph models that address only one formalism."],"supporting_citations":[{"why":"DGI establishes the mutual-information contrastive learning paradigm on graphs and provides the linear evaluation protocol CLDG adopts.","marker":"[13]"},{"why":"InfoNCE supplies the contrastive objective used in both the local and global CLDG losses.","marker":"[28]"},{"why":"GCN is the base encoder architecture used throughout the main CLDG experiments.","marker":"[20]"},{"why":"GRACE is a perturbation-based contrastive baseline that CLDG must beat to show the benefit of timespan views over corruptions.","marker":"[17]"},{"why":"CCA-SSG is a static contrastive baseline that motivates the comparison between feature-correlation pretraining and temporal contrastive pretraining.","marker":"[19]"},{"why":"DySAT is a discrete-time dynamic graph baseline that CLDG compares against on accuracy and on parameter and time efficiency.","marker":"[39]"},{"why":"TGAT is a continuous-time dynamic graph baseline that represents the sequential-model approach CLDG avoids.","marker":"[40]"},{"why":"MNCI is the strongest temporal-neighborhood baseline on several datasets and a key comparison for the global-level CLDG variant.","marker":"[58]"}],"fun_headline_variants":["No labels, no recurrence: CLDG wins on 7 graphs","Temporal views as free labels for dynamic graph learning","Each node's own past is its best teacher in graphs","Self-supervised graph learning from time shifts","CLDG: time as augmentation, 2000x lighter"],"cache_read_input_tokens":21760,"weakest_assumption_plain":"The premise is that a node's meaning and label stay stable across the time windows used as views, a property the paper observes empirically but does not quantify; the authors themselves note it fails when graph changes are non-continuous and labels constantly change.","fun_headline_variants_meta":{"raw":{"variants":["No labels, no recurrence: CLDG wins on 7 graphs","Temporal views as free labels for dynamic graph learning","Each node's own past is its best teacher in graphs","Self-supervised graph learning from time shifts","CLDG: time as augmentation, 2000x lighter"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000252,"raw_usage":{"total_tokens":1558,"prompt_tokens":938,"completion_tokens":620,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":554,"completion_tokens_details":{"reasoning_tokens":540}},"tokens_in":554,"tokens_out":620,"duration_ms":5324,"temperature":1.0,"reasoning_tokens":540,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T12:12:41.554620+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a dataset where node labels demonstrably change over time, such as a citation network re-labeled as researchers change fields or a synthetic temporal graph with periodic label flips. If CLDG's accuracy degrades toward or below a static contrastive baseline as the label-change rate rises, the temporal translation invariance assumption is the cause. A direct check is to compute same-node label agreement between non-overlapping windows and test whether it predicts CLDG's gain over perturbation-based baselines.","supporting_citations":[{"cited_title":"Deep graph infomax","cited_arxiv_id":null,"evidence_quote":"DGI establishes the mutual-information contrastive learning paradigm on graphs and provides the linear evaluation protocol CLDG adopts."},{"cited_title":"Representation learning with contrastive predictive coding,","cited_arxiv_id":null,"evidence_quote":"InfoNCE supplies the contrastive objective used in both the local and global CLDG losses."},{"cited_title":"From canonical correlation analysis to self-supervised graph neural networks,","cited_arxiv_id":null,"evidence_quote":"CCA-SSG is a static contrastive baseline that motivates the comparison between feature-correlation pretraining and temporal contrastive pretraining."},{"cited_title":"Inductive representation learning in temporal networks via mining neighborhood and community influences,","cited_arxiv_id":null,"evidence_quote":"MNCI is the strongest temporal-neighborhood baseline on several datasets and a key comparison for the global-level CLDG variant."}],"review_version":1}