{"id":"e455445a-1ef7-4673-b30e-1d72225f22c0","arxiv_id":"2505.17764","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A hub-aware random walk strategy, tuned per dataset by grid search, beats dynnode2vec on graph reconstruction for 8 of 9 temporal networks.","lead":"This paper introduces DeepHub, a dynamic graph embedding method that biases random walks according to node degree, and reports improved reconstruction F1 over dynnode2vec on 8 of 9 temporal networks. The result matters because it suggests hub structure is an overlooked factor in temporal graph representation learning.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Selection-bias flaw in evaluation: Dynnode2vec and DeepHub are both tuned and reported on last-snapshot reconstruction F1, so DeepHub's extra hyperparameters (degree-transform type plus p/u grid) may explain the 8/9 win instead of hub-awareness; a held-out evaluation is needed.","rationale":"The paper's strongest claim is narrow: on a per-method maximum last-snapshot F1 comparison, DeepHub beats dynnode2vec on 8/9 real temporal networks and reduces hub/non-hub F1 disparities in several. The method is clearly specified, the datasets are real and public, and averaging over 10 runs is a positive feature. However, every supporting quantity in Tables 2 and 5 is a maximum over a grid search computed on the same last snapshot that is then used for reporting. This is exactly the protocol under which a method with extra hyperparameters can manufacture an apparent win. DeepHub has at least one additional categorical axis (degree transform) and an additional probability u, so the comparison is not between two equally constrained hypotheses. The selected transform being always inverse or inverse-log is a post-selection observation, not a confirmation that hub-avoiding transitions are causally responsible. The reader's weakest assumption identifies this same selection-bias concern, and it is the most load-bearing threat to the central claim. A held-out-snapshot evaluation, ideally with equalized search budgets and confidence intervals, would settle the question directly. Until such an evaluation is run, the conditional verdict is appropriate.","tokens_in":9957,"tokens_out":4792,"duration_ms":45411,"concrete_test":"Hold out the final snapshot of each of the nine networks. Tune Dynnode2vec and DeepHub on snapshots 1..N-1 only (using the same grid style as §4), then compute reconstruction F1 on snapshot N exactly once. To equalize search freedom, also run a variant where DeepHub's degree-transform mode is fixed a priori to 'inverse' instead of being selected by the test metric. If DeepHub does not beat Dynnode2vec on a majority of the nine held-out snapshots, the 8/9 headline result is best explained by selection-on-the-report-metric rather than by hub-aware walks.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on comparing best last-snapshot graph-reconstruction F1 scores after per-method grid search (§4–§5). Dynnode2vec tunes p, q, and dimension; DeepHub tunes p, u, and the degree-transform type (normal/log/inverse/inverse-log, Table 5). In both cases the same last-snapshot F1 is used both as the selection criterion and as the reported outcome. This is selection on the test metric. Because DeepHub has at least one extra discrete axis (four transform modes) and an extra sampling probability u, its search space is larger than the Dynnode2vec p/q grid, and the observed advantage could reflect added tuning freedom rather than hub-aware sampling. The conclusion that inverse or inverse-log transforms are best is especially vulnerable: the transform type was selected per dataset by F1, so the fact that all selected types are inverse/inverse-log does not independently confirm that hub-avoiding walks improve embeddings. Several reported wins are also small (1.1%, 2.6%, 2.9%, 5.3%), and no confidence intervals or paired significance tests are reported for the per-method F1 differences. The reader's weakest assumption is therefore the load-bearing point: the empirical case for hub-awareness has not been separated from hyperparameter selection effects.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DeepHub, a dynamic graph embedding method that augments the dynnode2vec sampling procedure with degree-based transition scoring. In addition to the return parameter p and a uniform-move probability u, DeepHub chooses among four degree transforms (normal, log, inverse, inverse-log) to bias walks toward or away from hubs. The authors evaluate DeepHub and the dynnode2vec baseline on nine real temporal networks using micro-averaged graph reconstruction F1 on the last snapshot, and they also analyze hub/non-hub F1 differences with Spearman correlations and Mann-Whitney U tests. They report that DeepHub outperforms dynnode2vec on eight of nine datasets, and that the best per-dataset transform is always inverse or inverse-log, which they interpret as evidence that avoiding hubs improves embeddings.","tokens_in":10281,"tokens_out":4451,"duration_ms":35723,"significance":"The idea of making dynamic random-walk embeddings structurally aware of hubs is reasonable and not well explored, and the method is simple enough to be usable. The paper's strengths include a systematic comparison across nine publicly available datasets, a careful hub/non-hub fairness analysis, and explicit reporting of selected hyperparameters. However, the main empirical claim currently rests on an evaluation protocol in which the same last-snapshot reconstruction F1 is used both to select hyperparameters and to report final scores, and DeepHub has more tunable axes than the baseline. The significance of the paper will rise or fall on whether the 8-of-9 advantage survives an independent held-out evaluation.","major_comments":[{"comment":"The evaluation protocol tunes both dynnode2vec and DeepHub by maximizing the last-snapshot graph reconstruction F1 and then reports that same F1 as the outcome. Because DeepHub has additional hyperparameter axes (the uniform-move probability u and the four-way degree-transform choice) beyond dynnode2vec's p and q grid, the reported gains on eight of nine datasets could be produced by extra tuning freedom rather than by hub-aware sampling. The authors should separate model selection from evaluation, for example by using one or more held-out snapshots for selecting p, u, and transform type and reporting F1 only on a final test snapshot, or by reporting nested cross-validated estimates. Without such separation, the central claim is not established.","section":"Section 4, baseline tuning paragraph; Tables 2 and 5"},{"comment":"The conclusion that 'in all networks biased random walks tend to avoid hubs' is post hoc: the transform type (inverse or inverse-log) was chosen per dataset because it maximized the reported F1. This selection makes it unsurprising that inverse/inverse-log dominate the 'best' column, and it does not independently confirm that hub avoidance causes the improvement. A fair test would fix the transform family before seeing the evaluation data, or would compare hub-avoiding versus hub-favoring walks under identical selection protocols.","section":"Section 5, Table 5"},{"comment":"The abstract states that hub-aware walks 'improve downstream task performance,' but the experiments contain no downstream tasks (link prediction, node classification, or similar); the evaluation is exclusively intrinsic graph reconstruction. If downstream performance is claimed, the authors need to add such experiments; otherwise the claim should be restricted to graph reconstruction quality.","section":"Abstract and Section 6"},{"comment":"No standard deviations or confidence intervals are reported even though F1 is averaged over ten runs, and several claimed improvements are small (email-eu 1.07%, fb-forum 2.56%, college-msg 3.10%, ia-primary 2.86%). Without variance information or a paired significance test across runs, these individual wins are not distinguishable from noise. The authors should report per-configuration mean and standard deviation (or a paired test) for at least the final selected configurations.","section":"Section 4, ten-run averaging; Tables 2 and 5"}],"minor_comments":[{"comment":"The phrase 'minimal set satisfying the condition' is ambiguous; specify that H is minimal by cardinality (or by the described greedy construction), and note that there is a typo ('set od nodes').","section":"Definition 1"},{"comment":"The dataset named 'ia-primschool' in Table 1 is called 'ia-primary' in Tables 3-5; please unify the name.","section":"Table 1"},{"comment":"The assertion that CTDNE and STWalk 'exhibit noticeably higher graph reconstruction errors' is not supported by any table or figure; either report the numbers or omit the claim.","section":"Section 5, paragraph on CTDNE and STWalk"},{"comment":"The pseudocode uses 'N C' with a space and defines max_deg over neighbors, but it does not define behavior for isolated nodes or for walks of length one; a short clarification would improve reproducibility.","section":"Algorithms 1-3"},{"comment":"Adding error bars or at least the 10-run standard deviation to Figure 2 would help readers judge the small differences between methods.","section":"Figure 2"}],"recommendation":"major_revision","confidential_remarks":"The main concern is the selection-on-test evaluation protocol; this is fixable with held-out snapshots or nested validation. If the authors provide such evidence and report variability, I would support publication. The paper is within scope for the venue and the literature coverage is appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, the useful thing here is the hub-versus-non-hub F1 analysis across nine temporal networks, plus the straightforward extension of the same group's static hub-aware walk to dynnode2vec. The method is cheap, clearly described, and the results are consistent: 8 of 9 datasets show higher last-snapshot graph reconstruction F1, and the hub/non-hub F1 gap narrows on several datasets. That is worth knowing.\n\nThe soft spot is the one you'd expect from the protocol. Both DeepHub and the Dynnode2vec baseline are tuned by maximizing F1 on the last snapshot, and that same F1 is what gets reported. DeepHub adds extra axes to the search—p, u, and four degree-transform modes—so the observed advantage could be extra tuning freedom rather than hub awareness. The per-dataset selection of inverse or inverse-log transforms is the most exposed conclusion: the fact that all chosen transforms are inverse variants does not independently confirm that avoiding hubs helps, because the transform type was selected by the same F1. Several wins are also small (1.1%, 2.6%, 2.9%, 5.3%), and no confidence intervals or paired significance tests are reported, even though each configuration was averaged over 10 runs.\n\nThe abstract overclaims \"downstream task performance\"; there is no downstream task in the paper, only intrinsic graph reconstruction. That mismatch should be fixed. Also, no code is released, which limits quick checking, though the datasets are public. The citation pattern is fine—the self-citation to [20] is the direct prior work, not padding.\n\nI don't think these flaws are fatal. The paper is straightforward, the hub-bias mechanism is plausible, and the hub/non-hub disparity analysis is a useful empirical addition. But the central claim—that hub-aware walks, rather than the extra hyperparameter freedom, cause the improvement—has not been separated. That needs a held-out snapshot or nested tuning, plus variance reporting.\n\nThis is an incremental paper for the dynamic embedding subfield. It deserves a serious referee, but the referee should ask for a cleaner evaluation. My recommendation: conditional accept with revisions, not desk reject and not a strong reject.","headline":"A simple, transparent extension of hub-aware random walks to dynamic graphs, but the empirical claim is weakened by tuning on the test metric and the abstract overstates downstream gains.","tokens_in":10780,"tokens_out":2398,"would_cite":false,"duration_ms":25754,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68R10","05C82"],"pacs":[],"model":"deepseek-v4-flash","headline":"Hub-repelling random walks beat the standard baseline on 8 of 9 networks","keywords":["dynamic graph embedding","random walks","hubs","degree centrality","graph reconstruction","dynnode2vec","temporal networks","representation learning"],"falsifier":"Run DeepHub and dynnode2vec with a fixed, pre-registered hyperparameter budget and evaluate on a held-out snapshot or a downstream task such as link prediction; if DeepHub no longer beats dynnode2vec on most networks, the hub-awareness advantage claimed here does not transfer beyond the tuning procedure.","tokens_in":9834,"feed_emoji":"🕸️","tokens_out":4236,"duration_ms":29510,"temperature":0.7,"pith_summary":"This paper claims that dynamic graph embeddings built from random walks are distorted by hub overrepresentation, and that explicitly biasing walk transitions according to node degree corrects the distortion. To test this, the authors introduce DeepHub, a random walk strategy that scores neighbors by degree centrality with optional inverse or logarithmic scaling, and compare it against dynnode2vec on nine real temporal networks. DeepHub achieves higher graph reconstruction F1 on eight of the nine networks and reduces the gap between hubs and non-hubs in several cases. The paper's central claim is that hub-awareness is an overlooked factor in dynamic graph embedding, and that a lightweight change to walk sampling can improve intrinsic embedding quality across diverse networks.","feed_headline":"Hub-repelling random walks beat the standard baseline on 8 of 9 networks","feed_subtitle":"A degree-biased walk strategy raises reconstruction F1 for dynamic graph embeddings and narrows the hub gap.","key_machinery":"The core mechanism is a degree-biased transition distribution inside the random walk, governed by three parameters: backtracking probability p, uniform-move probability u, and a degree-scoring mode (normal, inverse, log, or inverse-log). Rather than handling hubness as a post-hoc alignment step, the bias is injected at sampling time within the standard dynnode2vec incremental-update loop, so the walk itself produces contexts in which less-connected nodes are not drowned out by high-degree neighbors.","core_discovery":"DeepHub modifies dynnode2vec's random walk sampling so that at each step, with probability p the walker backtracks to the previous node, with probability u it moves to a uniformly random neighbor, and otherwise it samples among neighbors with probabilities proportional to degree-based scores. The scores are either proportional to degree, or inversely proportional via 1 + max_deg - deg(n), optionally with a log transform, and in the configurations that worked best on the tested networks the walker avoids hubs. Across nine temporal networks, DeepHub's best-tuned F1 exceeds dynnode2vec's on eight of them, with relative improvements up to about 17.8 percent, and it lowers the hub versus non-hub F1 gap enough that a Mann-Whitney test can no longer reject equal distributions in three networks where dynnode2vec showed significant differences.","pith_inferences":["A natural testable extension is to combine hub-aware sampling with local-intrinsic-dimensionality-based walk biases, since both target the same failure mode of standard walks.","The inverse-log scaling that won most networks suggests an optimal trade-off curve between exploration and hub avoidance; tuning that curve per network may yield further gains beyond the grid used here.","If hub-awareness improves reconstruction F1, it likely also helps downstream tasks such as link prediction and node classification, but the paper does not test those tasks, so that transfer remains an open inference."],"forward_implications":["DeepHub improves graph reconstruction F1 over dynnode2vec on 8 of 9 real temporal networks, with the largest gains appearing on networks where hubs were strongly overrepresented.","The best configurations consistently use inverse or inverse-log degree scaling, meaning walkers that avoid hubs produce higher-quality embeddings in these networks.","Hub-aware sampling makes hub and non-hub node embeddings more comparable, with the hub versus non-hub F1 gap no longer statistically significant in three networks where it was significant for the baseline.","Because DeepHub changes only the sampling strategy, it can be dropped into the incremental-update framework of dynnode2vec without altering the model architecture or the embedding update step."],"supporting_citations":[{"why":"Supplies the baseline dynnode2vec method and the incremental-update framework that DeepHub extends.","marker":"[8]"},{"why":"Provides the node2vec random-walk embedding model that underlies the sampled-context training.","marker":"[4]"},{"why":"Provides DeepWalk, the uniform random walk baseline that motivates the uniform-move component of DeepHub.","marker":"[14]"},{"why":"Introduces the notion of hubs in high-dimensional data that motivates hub-aware sampling.","marker":"[15]"},{"why":"Defines the minimal-set hub criterion used to partition nodes into hubs and non-hubs in the evaluation.","marker":"[16]"},{"why":"Shows that hub-aware random walks improve static graph embeddings, the direct predecessor being extended to the dynamic setting.","marker":"[20]"},{"why":"Supplies the methodology for correlating node degree with embedding quality metrics in dynamic graphs.","marker":"[7]"}],"fun_headline_variants":["Hub-avoiding walks boost dynamic embedding F1 on 8 of 9 networks","DeepHub: dodge hubs in random walks, lift embedding performance","Rewiring random walks away from hubs improves temporal embeddings","Dynamic graph embeddings get a hub-awareness upgrade","Hub-dodging walks narrow the performance gap in embeddings"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's central evidence is a comparison where the same F1 metric on the last snapshot is used both to pick hyperparameters and to report results, so the reported gains may partly reflect DeepHub's extra tuning freedom rather than hub awareness itself.","fun_headline_variants_meta":{"raw":{"variants":["Hub-avoiding walks boost dynamic embedding F1 on 8 of 9 networks","DeepHub: dodge hubs in random walks, lift embedding performance","Rewiring random walks away from hubs improves temporal embeddings","Dynamic graph embeddings get a hub-awareness upgrade","Hub-dodging walks narrow the performance gap in embeddings"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000172,"raw_usage":{"total_tokens":1259,"prompt_tokens":910,"completion_tokens":349,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":526,"completion_tokens_details":{"reasoning_tokens":264}},"tokens_in":526,"tokens_out":349,"duration_ms":3829,"temperature":1.0,"reasoning_tokens":264,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:40:45.935369+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run DeepHub and dynnode2vec with a fixed, pre-registered hyperparameter budget and evaluate on a held-out snapshot or a downstream task such as link prediction; if DeepHub no longer beats dynnode2vec on most networks, the hub-awareness advantage claimed here does not transfer beyond the tuning procedure.","supporting_citations":[{"cited_title":"In: 2018 IEEE international conference on big data (Big Data)","cited_arxiv_id":null,"evidence_quote":"Supplies the baseline dynnode2vec method and the incremental-update framework that DeepHub extends."},{"cited_title":"Journal of Machine Learning Research 11(sept), 2487–2531 (2010)","cited_arxiv_id":null,"evidence_quote":"Introduces the notion of hubs in high-dimensional data that motivates hub-aware sampling."},{"cited_title":"Computing99, 1055–1079 (2017)","cited_arxiv_id":null,"evidence_quote":"Defines the minimal-set hub criterion used to partition nodes into hubs and non-hubs in the evaluation."},{"cited_title":"Statistical Analysis and Data Mining: The ASA Data Science Journal 17(2), e11676 (2024)","cited_arxiv_id":null,"evidence_quote":"Shows that hub-aware random walks improve static graph embeddings, the direct predecessor being extended to the dynamic setting."},{"cited_title":"In: Cherifi, H., Donduran, M., Rocha, L.M., Cherifi, C., Varol, O","cited_arxiv_id":null,"evidence_quote":"Supplies the methodology for correlating node degree with embedding quality metrics in dynamic graphs."}],"review_version":1}