{"id":"fa78d6c8-03d5-48ee-a692-d889519b1619","arxiv_id":"1908.08402","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"TNA, a stacked GCN-GRU model with variational sampling, predicts new edges in temporal graphs and outperforms baselines on three real datasets.","lead":"This paper introduces Temporal Neighbourhood Aggregation, a model that combines graph convolutions with recurrent units to predict future links in evolving graphs. The authors report gains of up to 23% over baselines on real-world temporal networks while using no learnable decoder parameters.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Fixed-vertex-set handling is unspecified: identity features plus shared GCN weights require a single |V| per sequence, while Wiki/UCI have changing vertex sets; the 23% new-edge gain is unverifiable without the missing preprocessing.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern: the model assumes a fixed, consistently ordered vertex set while the chosen real datasets contain vertex arrivals, and the paper never describes the reconciliation. I considered two other candidate concerns. The parameter-count contradiction between the abstract and Table III is real: TNA reports 133K parameters versus 120-125K for GAE, GVAE, TO-GAE, TO-GVAE, and D-GCN on Bitcoina, so 'also requiring fewer overall model parameters' is not true relative to several named baselines. However, that weakens a supporting clause of the claim without undermining the performance comparison. The absence of negative-sampling details for AUC/AP is also important, but it affects baselines and proposed model jointly and is less clearly a structural flaw. The fixed-vertex-set issue cuts deeper: it touches the input representation (identity features), the architecture (shared GCN weights in Eq. 3), the variational objective (Eq. 4), and the evaluation (Algorithm 1). If the implementation restricts to common vertices, then the headline task is no longer new edges in the original dataset but new edges among persistent vertices. If it pads to a union, Eq. 4 and the training procedure are inaccurate. The paper's own future-work sentence in Section VI, 'enabling vertex arrival to be modelled,' is an internal admission that the current model does not handle vertex arrival. Because the code is open-source, this is checkable rather than fatal, so the appropriate verdict remains CONDITIONAL as the reader concluded. I agree with the reader and do not recommend changing the verdict.","tokens_in":15809,"tokens_out":5336,"duration_ms":61530,"concrete_test":"Inspect the released repository (github.com/sbonner0/temporal-neighbourhood-aggregation) on Wiki-Vote and UCI. In the data-loading path, record |V_t| for each snapshot and compare with the size of the union vertex set; compute the fraction of Et\\Et-1 edges whose endpoints first appear at time t. Then identify the code path that builds X and A: if it uses the union size with zero rows for absent vertices, the fixed-vertex assumption is imposed and the paper should state that; if it truncates to V_t, shared W_g is impossible and Table III cannot be reproduced as written. Finally, rerun the Table III TNA row under explicit common-vertices-only preprocessing and measure whether the 23% new-edge advantage persists; if the margin shrinks materially, the abstract's claim applies to a restricted subset of the named datasets.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is empirical superiority in predicting new temporal edges. That claim rests on the model being well-defined on the datasets in Table II. In Section III-C, all TNA parameters are described as shared through time, and Table I sets X to the identity matrix, so the first GCN weight matrix W_g^(1) in Eq. (1) has input dimension |V|. Equation (3) applies the same TNA block to every snapshot. If |V_t| changes across t, this shared weight matrix cannot be defined. Equation (4) nevertheless indexes the variational distribution by |V_t|, so the paper itself allows time-varying vertex sets without reconciling them with parameter sharing. Wiki-Vote and UCI are networks where vertices first appear after the initial snapshot; Section VI confirms this by listing 'enabling vertex arrival to be modelled' as future work. Algorithm 1 has no preprocessing step that constructs a common vertex universe or excludes appearing vertices. Therefore a run of the method must make an unstated choice: (a) truncate to vertices present in all snapshots, dropping new-edge events that involve appearing vertices and changing the AUC/AP test set; (b) pad all union vertices as zero rows, contradicting Eq. (4) and potentially leaking future vertex existence into earlier snapshots; or (c) retrain per-snapshot sizes, breaking the shared-parameter claim. Without knowing which choice was made, the reported margins over D-GCN and the Dyn* baselines are not reproducible from the text alone.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Temporal Neighbourhood Aggregation (TNA), a temporal graph embedding architecture that stacks GCN layers with GRU recurrence at each hop, adds layer normalization and skip connections, and produces vertex representations via variational sampling. The model is trained end-to-end to predict the next temporal snapshot using an inner-product decoder with zero learnable parameters. The authors evaluate TNA on three real-world datasets (Bitcoin-Alpha, Wiki-Vote, UCI-Messages) and two synthetic datasets, comparing it against static and temporal baselines (GAE, GVAE, TO-GAE, TO-GVAE, DynAE, DynRNN, DynAERNN, D-GCN). The central claim is that TNA outperforms competing methods at predicting new temporal edges by up to 23% on real datasets while using fewer model parameters. The paper also includes ablation studies showing the contribution of each architectural component.","tokens_in":16089,"tokens_out":2551,"duration_ms":27068,"significance":"If the reported results hold under a clearly specified and reproducible evaluation protocol, TNA would be a useful parameter-efficient baseline for temporal link prediction, with the interesting design choice of placing recurrence at each neighborhood depth rather than only at the graph level. The paper ships a public code repository, and the ablation study in Table IV is a useful contribution. However, the significance of the empirical claims is currently limited by an underspecified evaluation protocol, and the model's ability to handle time-varying vertex sets is not established. The core new-edge prediction claim therefore cannot yet be assessed as reliable.","major_comments":[{"comment":"The model as specified assumes a fixed, consistently ordered vertex set. Table I sets X to the identity matrix of A, so the first GCN weight matrix in Eq. (1) has input dimension |V|, and Eq. (3) applies the same shared-parameter TNA block to every snapshot. However, Eq. (4) indexes the variational distribution by |V_t|, and the datasets Wiki and UCI contain vertices that appear after the initial snapshots. Section VI explicitly lists 'enabling vertex arrival to be modelled' as future work, confirming that the current model does not handle arrival. Algorithm 1 contains no preprocessing step that constructs a common vertex universe, trims appearing vertices, or otherwise reconciles time-varying vertex sets with shared weights. The authors must specify exactly how this was handled; otherwise the reported AUC/AP numbers are not reproducible, and options such as padding with zero rows could leak information about future vertex existence into earlier snapshots.","section":"Section III-C, Table I, Equations (1) and (4); Section VI"},{"comment":"The evaluation protocol for new-edge prediction is underspecified with respect to negative sampling. The task is defined as predicting edges in E_t \\ E_{t-1}, and AUC and AP are computed, but the paper never states how negative examples are sampled for these metrics. This is load-bearing because AUC and AP are ranking measures over a positive set and a negative set; different negative-sampling strategies (e.g., random non-edges, same-degree non-edges, or non-edges with geographic constraints) can substantially change the reported scores and the comparison with baselines. The authors should describe the negative sampling distribution, the ratio of negatives to positives, and whether the same negative set is used for all methods.","section":"Section IV-A, Algorithm 1, Section V-B"},{"comment":"The meaning of the '25%, 50%, 100%' columns in Table III is ambiguous and the text does not define it. Algorithm 1 describes iterating over all G_t with t ≥ 3 and training on G_1,...,G_{t-1}, but Table III reports results 'when predicting at various percentages of the length of the time-sequence.' It is unclear whether the models are trained on only the first 25%/50% of snapshots and then evaluated on all later snapshots, or whether each test graph is predicted using only the immediately preceding fraction of history. These different protocols lead to different difficulties and different numbers. The authors should define the training/test split precisely and relate it to Algorithm 1.","section":"Table III, Section V-B"},{"comment":"The paper states that TNA 'significantly outperforms' the baselines, but no statistical significance tests are reported. The results are presented as means with standard deviations over (apparently) three random seeds, and for several cells the standard deviations overlap substantially (e.g., Wiki AUC at 100%: TNA 0.634 ± 0.050 vs D-GCN 0.563 ± 0.087, where the overlap is not huge but exists; Bitcoina AUC at 25%: TNA 0.665 ± 0.067 vs D-GCN 0.622 ± 0.084). At minimum, the authors should state the number of runs and report paired significance tests (or confidence intervals) for the headline comparisons, and they should avoid the word 'significant' unless such tests support it.","section":"Section V-B, Table III"}],"minor_comments":[{"comment":"There are numerous typos and slight formatting issues, including 'scientific', 'varitonal', 'Iternational', 'disregard' (in the conclusion), and inconsistent capitalization of 'Variational' in Table IV; these should be corrected.","section":"Throughout"},{"comment":"The D-GCN baseline is described only briefly as 'a dynamic GCN similar to approaches in [20] and [21].' Since D-GCN is a primary competitor in Table III and Figure 4, the authors should provide more implementation detail (depth, hidden sizes, training schedule) or refer to a public implementation.","section":"Section IV-D"},{"comment":"For the SBM dataset, TNA's AUC and AP are reported as 0.502 ± 0.024 for both metrics, which is suspiciously identical and likely a typo; the same applies to the R-Cora row. The authors should check these entries.","section":"Table V"},{"comment":"The paragraph describing the training procedure mentions 'full-batch gradient descent with RMSProp,' but RMSProp is a stochastic optimizer; 'full-batch' is ambiguous and should be clarified (e.g., whether each update uses all vertices in the snapshot).","section":"Section III-E"},{"comment":"The dataset table reports 'First Edge' and 'Last Edge' dates, but the text says Wiki-Vote covers '28th March 2004' while the table says '28/02/2005'; this inconsistency should be fixed.","section":"Section IV-C, Table II"}],"recommendation":"major_revision","confidential_remarks":"The core architectural idea is reasonable and the ablation study is informative, but the evaluation protocol (vertex set handling, negative sampling, and the history percentage convention) is not sufficiently specified to support the claimed margin over baselines. The good news is that these are fixable with a clear description and, ideally, a released evaluation script. I would give the authors a major-revision opportunity rather than rejecting, but the burden is on them to show that the reported improvements are not an artifact of an unstated preprocessing or sampling choice."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core contribution is real: TNA stacks GCN, GRU, and variational sampling with a skip connection to capture temporal change at multiple neighborhood depths, and does so without a learned decoder. That is a clean, useful combination, and the ablation in Table IV gives proper credit to each component. The main empirical claim also mostly holds: on the three real datasets, TNA beats D-GCN and the other baselines on new-edge prediction in Table III, often by a wide margin on Bitcoin and UCI. The code is claimed public, which is good.\n\nThe soft spots are in the evaluation, not the architecture. First, the abstract says TNA requires fewer parameters, but Table III shows it has more parameters than GAE, GVAE, TO-GAE, TO-GVAE, and D-GCN on all three datasets. It is only parameter-lean relative to the Dyn* family. That claim should be revised.\n\nSecond, the stress-test concern is legitimate. The model uses identity features and shared weights, which presupposes a fixed vertex set, but Wiki-Vote and UCI have vertices appearing over time. The paper lists vertex arrival as future work and says no preprocessing is needed, yet Algorithm 1 says pre-process. An unstated decision about how the vertex set is fixed affects which new edges are even testable. The public code may resolve this, but the paper alone does not make the 23% margin reproducible.\n\nThird, the evaluation protocol is underspecified. There is no description of how negative edges are sampled for the new-edge task, no significance tests, and the 25/50/100% history convention is ambiguous. On Wiki, the AUC gap over D-GCN is within one standard deviation at 25% and 50%, so 'superior' should be tempered. The synthetic results also show TNA not beating baselines on SBM and R-Cora, which the paper glosses over.\n\nThe math is standard and the citation pattern is fine. Self-citation of TO-GAE/TO-GVAE is not a problem; those are reasonable baselines, and D-GCN is a fair stronger comparison.\n\nWho is this for? Anyone working on temporal link prediction or snapshot-based graph embeddings. It deserves a serious referee, but the revision needs to document preprocessing, negative sampling, and significance, and fix the parameter-count claim. If the code checks out, the empirical contribution is solid.","headline":"A genuinely new architecture with strong real-data results, but the evaluation write-up has documentation gaps that need fixing before the headline numbers can be trusted.","tokens_in":664,"tokens_out":1731,"would_cite":true,"duration_ms":42304,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A temporal graph model, Temporal Neighbourhood Aggregation, predicts new edges in future graph snapshots up to 23% more accurately than competing deep baselines while using fewer parameters.","keywords":["temporal graph representation learning","dynamic link prediction","graph convolutional networks","recurrent neural networks","variational inference","next-graph prediction","temporal neighbourhood aggregation"],"falsifier":"Run TNA on the Wiki-Vote or UCI-Messages snapshots exactly as Algorithm 1 is written, without adding a vertex-alignment step; if the fixed $|V|\\times|V|$ inputs and the new-edge evaluation are undefined or the reported AUC/AP cannot be reproduced, the claim that the model needs no preprocessing is false.","tokens_in":15591,"feed_emoji":"🔗","tokens_out":8964,"duration_ms":81657,"temperature":0.7,"pith_summary":"The paper introduces Temporal Neighbourhood Aggregation (TNA), a vertex-embedding model that reads a series of graph snapshots and tries to predict the next graph in the series. Each TNA block combines a graph-convolution layer with a gated recurrent unit, and stacking two blocks lets a vertex draw on temporal change in both its one-hop and two-hop neighbourhoods. No vertex features, labels, or random walks are required. The final embeddings are sampled from a Gaussian using variational inference, and the next adjacency matrix is decoded by a parameter-free inner product. On three real-world temporal networks the model predicts previously unseen edges more accurately than the compared deep baselines, with the paper reporting gains of up to 23% on real-world datasets while using fewer parameters than the recurrent baseline family.","feed_headline":"Temporal graph model beats deep baselines by up to 23%","feed_subtitle":"It combines graph convolutions with recurrence at multiple depths to predict future links with fewer parameters.","key_machinery":"The central object is the TNA block: one graph-convolution layer followed by a gated recurrent unit, with a concatenated skip connection and layer normalisation. In the block, the GCN aggregates a vertex's current one-hop neighbourhood, the GRU reads the resulting vertex representations across snapshots so the block carries memory of how that neighbourhood has been changing, and the linear mixing layer learns how much weight to give topology versus time at each depth. Stacking two TNA blocks aggregates temporal information from two hops, and the variational sampling head, a pair of GCN layers parameterising mean and variance with the reparameterisation trick, produces the final embedding. The generative decoder is simply the sigmoid inner product of embedding pairs, which is why the model has no parameterised decoder and only 133K–239K parameters on the test graphs.","core_discovery":"TNA's central claim is that a temporally recurrent graph convolution can capture how each vertex's neighbourhood changes at several depths and convert that directly into an accurate forecast of the next graph. The model computes, for each snapshot, a graph-convolved representation, passes it through a GRU that carries state across snapshots, and mixes the two with a learned linear combination after layer normalisation; two such blocks stacked in depth give each vertex access to two hops of temporal neighbourhood change. Variational sampling then produces the embedding matrix $Z_t$ from two GCN heads that estimate mean and variance, and the future graph is reconstructed as a sigmoid inner product of $Z_t$ with itself, so the decoder has no trainable parameters. The training objective is the variational lower bound on $\\log p(A_{t+1}|Z_t)$ minus the KL divergence to a Gaussian prior, plus a small L2 penalty. In the paper's experiments TNA attains the highest mean AUC and average precision for new-edge prediction on Bitcoin-Alpha, Wiki-Vote, and UCI Messages at every tested fraction of the time history, with 133K–239K parameters, while the DynAE/DynRNN/DynAERNN baselines use millions.","pith_inferences":["As an extension, we would expect the same architecture to need an explicit vertex-alignment or inductive aggregation step on graphs where vertices appear and disappear, because the fixed $|V|\\times|V|$ input and identity features are only well-defined for a stable vertex set.","We would also test how sensitive the 23% gain is to snapshot granularity; monthly batching of Bitcoin-Alpha versus weekly batching of UCI changes how much edge evolution each recurrent step sees, and the model's advantage over D-GCN may shrink when snapshots are so fine that few new edges appear per step.","A natural component study to run elsewhere is the full ablation from the paper's Table IV on Wiki-Vote and UCI; if variational sampling or the skip connection contributes less on those graphs, the one-architecture-fits-all claim would need qualification.","A testable variant would swap the fixed-graph convolution for an aggregator that handles previously unseen vertices; that would separate the contribution of temporal recurrence from the contribution of spectral convolution."],"forward_implications":["Because the decoder is parameter-free, total model size is set by the encoder alone, which is what allows the method to scale to graphs of thousands of vertices and long snapshot histories without the memory cost of recurrent decoders.","The ablation shows that including second-hop temporal information raises AUC and AP by roughly 10 percentage points over first-hop-only on Bitcoin-Alpha, so multi-depth recurrence is the component that carries most of the predictive gain.","Since no features, labels, or random walks are needed, the model applies directly to any snapshot graph sequence, not just graphs with rich node attributes.","The multi-step forecasting experiment indicates that feeding the model's own predicted graph back in as input keeps new-edge prediction accuracy above that of the closest baseline for several steps ahead.","On full-graph reconstruction, TNA remains competitive on Bitcoin-Alpha and Wiki-Vote and slightly behind on UCI, showing the gain on new edges does not come from ignoring existing edges."],"supporting_citations":[{"why":"Supplies the spectral graph-convolution layer that forms the topological half of each TNA block.","marker":"[1]"},{"why":"Supplies the variational graph auto-encoder framework, including the inner-product decoder and the reconstruction objective used for link prediction.","marker":"[12]"},{"why":"Defines the DynAE, DynRNN, and DynAERNN baselines whose large decoders and parameter counts are the main comparison against TNA.","marker":"[19]"},{"why":"Provides the reparameterisation trick and the variational lower-bound objective used to sample and train the final vertex embeddings.","marker":"[33]"},{"why":"Defines the temporal offset reconstruction procedure used by the TO-GAE and TO-GVAE baselines.","marker":"[34]"},{"why":"Supplies the layer-normalisation operation that stabilises training inside each TNA block.","marker":"[31]"},{"why":"Provides the Bitcoin-Alpha and Wiki-Vote temporal graph datasets used in the real-world evaluation.","marker":"[35]"},{"why":"Provides the UCI Messages temporal graph dataset used in the real-world evaluation.","marker":"[36]"}],"fun_headline_variants":["Temporal graph forecasting leaps 23% via recurrent variational GCNs","Temporal Neighbourhood Aggregation predicts links 23% better","Recurrent variational graph convolutions beat rivals by 23% on temporal links","TNA model: 23% better temporal link prediction with fewer parameters","Predict future graph links 23% more accurately with temporal recurrence"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pipeline assumes the vertex set is fixed and identically ordered across all snapshots, so a single $|V|\\times|V|$ adjacency matrix and identity feature matrix can represent every graph.","fun_headline_variants_meta":{"raw":{"variants":["Temporal graph forecasting leaps 23% via recurrent variational GCNs","Temporal Neighbourhood Aggregation predicts links 23% better","Recurrent variational graph convolutions beat rivals by 23% on temporal links","TNA model: 23% better temporal link prediction with fewer parameters","Predict future graph links 23% more accurately with temporal recurrence"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000619,"raw_usage":{"total_tokens":2888,"prompt_tokens":980,"completion_tokens":1908,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":596,"completion_tokens_details":{"reasoning_tokens":1814}},"tokens_in":596,"tokens_out":1908,"duration_ms":104127,"temperature":1.0,"reasoning_tokens":1814,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:55:56.613536+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run TNA on the Wiki-Vote or UCI-Messages snapshots exactly as Algorithm 1 is written, without adding a vertex-alignment step; if the fixed $|V|\\times|V|$ inputs and the new-edge evaluation are undefined or the reported AUC/AP cannot be reproduced, the claim that the model needs no preprocessing is false.","supporting_citations":[{"cited_title":"Semi-supervised classification with graph convolutional networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the spectral graph-convolution layer that forms the topological half of each TNA block."},{"cited_title":"dyngraph2vec: Captur- ing network dynamics using dynamic graph representation learning,","cited_arxiv_id":null,"evidence_quote":"Defines the DynAE, DynRNN, and DynAERNN baselines whose large decoders and parameter counts are the main comparison against TNA."},{"cited_title":"Temporal graph offset reconstruction: Towards tempo- rally robust graph representation learning,","cited_arxiv_id":null,"evidence_quote":"Defines the temporal offset reconstruction procedure used by the TO-GAE and TO-GVAE baselines."},{"cited_title":"Konect: the koblenz network collection,","cited_arxiv_id":null,"evidence_quote":"Provides the UCI Messages temporal graph dataset used in the real-world evaluation."}],"review_version":1}