{"id":"637a7949-67eb-42c7-a6fc-871a9c3c0ee5","arxiv_id":"2507.07595","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Context Pooling improves inductive link prediction in knowledge graphs by building a query-specific subgraph that keeps only neighbors whose relation types co-occur with the query relation.","lead":"A new method called Context Pooling selects only the neighbors whose relation types are statistically associated with the query relation, leaving out irrelevant and noisy neighbors. Adding this selection layer to two leading graph neural network link predictors improves accuracy on most of the standard transductive and inductive knowledge graph benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 42/48 SOTA claim is unverified because baseline numbers in Tables 1–2 may come from a different negative-sampling protocol than the one the paper applies to its own models.","rationale":"The reader's weakest assumption (Assumption 1 underlying the optimized Algorithm 3) is a legitimate and well-documented approximation, and the ablation in Table 4 indeed shows a measurable performance gap on FB15k-237-V1. However, that concern does not directly invalidate the paper's empirical efficacy claim: the optimized version still improves over the RED-GNN/NBFNet baselines in most settings, so the method could be a useful heuristic even if the independence assumption is false. The more load-bearing issue is the comparability of the SOTA evaluation itself. The abstract's central quantitative claim is 'SOTA performance in 42 out of 48 settings,' and this claim is computed against baseline numbers that the paper does not explicitly say were obtained under the same test protocol. Because Section 5 describes adopting the all-entity-negative protocol from RED-GNN, but never states that Neural LP, DRUM, CompGCN, AnyBURL, and GRAIL were re-run under that protocol, there is a real risk that those baseline values came from the originally published 49-negative setup, which is easier and inflates their scores. If so, RED-GNN+CP and NBFNet+CP would appear to beat baselines that were evaluated under a different, easier ranking condition, making the 'SOTA in 42/48' statement an artifact of inconsistent evaluation rather than a genuine improvement. This is a testable, concrete concern, and it is more central to the headline claim than the independence approximation, since it affects every reported SOTA cell rather than just the optimized algorithm's fidelity. I therefore keep the reader's conditional verdict: the paper needs either a clear statement plus evidence that all baselines were run under the stated protocol, or corrected baseline tables. My agreement is partial because I recognize the reader's independence-assumption concern as valid but see the evaluation-protocol issue as the single most load-bearing for the paper's central claim.","tokens_in":15597,"tokens_out":19630,"duration_ms":200536,"concrete_test":"Re-run every baseline model (Neural LP, DRUM, CompGCN, AnyBURL, GRAIL, NBFNet, RED-GNN) under the all-entities-as-negatives protocol described in Section 5, using the authors' released code or original implementations. Then recompute the number of settings (out of 48) where NBFNet+CP or RED-GNN+CP is the best. If the baseline numbers decrease relative to Tables 1–2 (e.g., any MRR or Hit@1 value drops by more than 5%), the SOTA count will change and the headline claim must be revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 5 states that, unlike the setup of GRAIL [21] which samples 49 random negatives per query, the paper uses the all-entities-as-negatives protocol from RED-GNN [31]. Yet the paper never states that every baseline row in Tables 1 and 2 was recomputed under this same protocol. The baseline values for Neural LP, DRUM, CompGCN, AnyBURL, and GRAIL are widely reported in prior papers using the easier 49-negative setting, which makes ranking tasks easier and inflates MRR/Hits@1. If those rows were copied from prior publications, then comparing them to RED-GNN+CP/NBFNet+CP numbers computed under the more challenging all-entities protocol is not apples-to-apples. The headline claim 'SOTA performance in 42 out of 48 settings' depends entirely on this comparison, but the paired t-test in Section 5.1.3 only validates the improvement of CP over NBFNet and RED-GNN, not over the other baselines. This is a load-bearing gap in the central claim: the paper asserts SOTA without demonstrating that the SOTA comparison is fair.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Context Pooling (CP), a training-free, query-specific graph pooling method for inductive knowledge-graph link prediction. CP scores neighboring relations by empirical neighborhood precision and recall computed from the training graph, prunes neighbors that fall below thresholds, and feeds the resulting query-specific subgraph alongside the original graph into a GNN such as NBFNet or RED-GNN. The authors provide an exact power-set formulation (Algorithm 1), an optimized formulation (Algorithm 3) under a conditional-independence assumption, a proof that the optimized relevance scores are proportional to products of singleton scores, and experiments on three datasets with four versions each in transductive and inductive settings. They report that CP improves the two base models and achieves SOTA performance in 42 of 48 settings.","tokens_in":15877,"tokens_out":9565,"duration_ms":95518,"significance":"If the SOTA claim can be supported under a uniform evaluation protocol, CP is a simple and generic plug-in component with inductive capability, and the paper would make a useful contribution to KG link prediction. Strengths include the clear formalization of Definitions 5 and 6, the explicit independence assumption rather than a hidden one, the released code, and the paired t-test evidence that CP improves NBFNet and RED-GNN. The co-occurrence statistics are computed from the training graph rather than from test labels, so the method is not circular. The main reservations are the fairness of the baseline comparison and the limited validation of the independence assumption used by the optimized algorithm.","major_comments":[{"comment":"The SOTA comparison may not be apples-to-apples. Section 5 states that the authors use the all-entities-as-negatives protocol from RED-GNN [31], unlike GRAIL's 49-negative protocol, but the manuscript never states that the rows for Neural LP, DRUM, CompGCN, AnyBURL, and GRAIL were recomputed under this protocol. If those rows are taken from prior papers that used an easier negative-sampling setup, their MRR and Hits@1 values are inflated relative to the CP models, and the headline \"42 out of 48 settings\" is not established. The paired t-test in Section 5.1.3 only compares CP against NBFNet and RED-GNN, so it does not validate superiority over the other baselines. Please recompute all baselines under the same protocol or explicitly restrict the SOTA claim to a fair comparison.","section":"Section 5, Tables 1-2"},{"comment":"Assumption 1 is load-bearing and only weakly validated. All main results use Algorithm 3, whose correctness rests on the conditional-independence assumption. The ablation in Section 5.3 (Table 4) is the only evidence for this assumption, and it is mixed: on FB15k-237-V1 the unoptimized Algorithm 1 gives MRR 0.396 vs 0.383 for the optimized algorithm (+3.4%) and Hit@1 0.332 vs 0.316 (+5.1%), while on WN18RR-V1 the optimized version is slightly better. The paper attributes the WN18RR-V1 result to noise, but the FB15k-237-V1 gap shows that the independence assumption can change the selected neighbors and measurably reduce accuracy. Moreover, the unoptimized algorithm in the ablation is not exactly Algorithm 1, because Section 5.3 restricts the relation count to [4,6). Please validate the assumption on more splits/datasets, report results for Algorithm 1 where feasible, or provide a theoretical bound on the error introduced by the approximation.","section":"Section 4.3, Table 4"},{"comment":"The scoring function and thresholds are under-specified. Definition 4 defines CNF via a generic Rel(), and Definitions 5-6 define two separate metrics, neighborhood precision and neighborhood recall. Algorithm 3 line 12 tests Rel(r', r_i) > R0' but the paper never states whether the threshold is applied to precision, to recall, or to a combination (e.g., product or minimum), nor how the two thresholds are chosen. Section 5 only says the thresholds \"are set within the range [10^-5, 10^-1]\", which spans four orders of magnitude and is not a reproducible specification. Please state the exact scoring rule and per-dataset threshold values, or the criterion used to select them.","section":"Section 4.3, Algorithm 3, Section 5"}],"minor_comments":[{"comment":"The counts in the text do not match the tables: after stating RED-GNN+CP is best in 10/12 transductive settings, the text says \"In the remaining 3 of 4 settings\" instead of 2, and after 8/12 inductive settings it says \"remaining 5 of 8 settings\" instead of 4. Please correct these counts.","section":"Sections 5.1.1 and 5.1.2"},{"comment":"The novelty claim that Context Pooling is the first graph pooling method for KGs would be stronger with a brief comparison to prior work on selectively aggregating neighbors in knowledge graphs; currently the related work only covers homogeneous-graph pooling.","section":"Section 2.2 / Section 1"},{"comment":"In Table 3, the '-1' suffix used for inverse relations is not explained in the caption; please clarify the notation and specify whether the precision and recall columns are separate neighbor sets or a combined set.","section":"Table 3"},{"comment":"The phrase \"Markov chain-inspired approach\" is misleading because Theorem 1 follows directly from the independence assumption and does not require Markov chain machinery; consider removing the reference or rewording.","section":"Section 4.3"}],"recommendation":"major_revision","confidential_remarks":"The paper has a solid core idea and the relative gains over NBFNet and RED-GNN are plausible, but the experimental section needs careful reworking: the baseline protocol must be clarified or the bases recomputed, and the independence assumption behind the optimized algorithm needs stronger validation. The numerical inconsistencies in the result counts also suggest a careful pass over the tables. These issues are addressable within the scope of a revision, so I do not recommend rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the pooling idea is genuinely new and the integration with NBFNet/RED-GNN looks clean, but the headline SOTA claim rests on a comparison that may not be apples-to-apples. The paper never states that every baseline row in Tables 1–2 was recomputed under the all-entities negative-sampling protocol it uses for its own models. If those rows come from earlier papers that used the easier 49-negative setup, the comparison is unfair and the '42/48 settings' claim is unsupported.\n\nWhat's actually new: the neighborhood precision/recall metrics defined at the relation level, and the query-specific graph construction using those scores. This is a sensible, training-free way to prune neighbors, and it plugs into GNN link predictors as a generic layer. The math in Theorem 1 is correct under Assumption 1, and the ablation honestly shows that the independence assumption costs a bit of accuracy on FB15k-237-V1 (0.396 vs 0.383 MRR) while saving a lot of time. Code is public.\n\nSoft spots: (1) The negative sampling protocol problem. The paper explicitly says it uses the all-entities protocol from RED-GNN, but lists baseline rows (Neural LP, DRUM, CompGCN, AnyBURL, GRAIL) without stating they were recomputed. If those numbers are copied from originals that used 49 negatives, the comparison is not valid. (2) The 42/48 claim doesn't align with the text: Section 5.1.1 reports 10/12 MRR and 10/12 Hit@1 wins for RED-GNN+CP, and inductive 8/12 each, then mentions NBFNet+CP winning some others, but never cleanly counts 42. The abstract's number looks inflated or miscounted. (3) Thresholds are only given as a range [1e-5, 1e-1]; exact values aren't reported, which hurts reproducibility. (4) The 'first graph pooling in KGs' claim needs more care—NBFNet and RED-GNN already build query-specific subgraphs, so the novelty is the pooling criterion, not the query-specific graph per se.\n\nOverall: a solid, practical idea with a credible evaluation setup except for the baseline issue. The core method is worth knowing about. I'd send it to review but require the authors to clarify the baseline protocol and fix the claimed win count.","headline":"A genuinely new pooling layer for KG link prediction with clean integration, but the SOTA claim is shaky because the baseline protocol isn't confirmed and the win count doesn't add up.","tokens_in":16324,"tokens_out":3205,"would_cite":true,"duration_ms":31115,"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":"Context Pooling claims that query-specific graph pooling based on neighborhood precision and recall improves inductive link prediction in knowledge graphs, achieving state-of-the-art results in 42 of 48 settings.","keywords":["Knowledge Graphs","Link Prediction","Graph Neural Networks","Graph Pooling","Inductive Link Prediction","Query-specific Graphs","Neighborhood Precision","Context Neighbor Family"],"falsifier":"Construct or select a KG where the co-occurrence of neighboring relations is strongly correlated with the query relation, for instance a domain where the relation Profession usually co-occurs with Award only when Company is also present, and run RED-GNN+CP with Algorithm 3 versus plain RED-GNN. If the pooled model's MRR drops below the base model while the unoptimized Algorithm 1 does not drop, the independence assumption is the load-bearing approximation.","tokens_in":15418,"feed_emoji":"🔗","tokens_out":7186,"duration_ms":72691,"temperature":0.7,"pith_summary":"Context Pooling attempts to establish that GNN-based link prediction in knowledge graphs is better when the model does not aggregate all neighbors, but only neighbors judged logically relevant to the query relation. The paper defines two statistics, neighborhood precision and neighborhood recall, computed from relation co-occurrence in the training graph, and uses them to build a query-specific context graph at every aggregation hop. The method is generic: wrapping RED-GNN and NBFNet with Context Pooling produces state-of-the-art results in 42 of 48 transductive and inductive settings across WN18RR, FB15k-237, and NELL-995. If the claim holds, graph pooling becomes a cheap, training-free plug-in that directly targets the recent observation that vanilla neighbor aggregation contributes little to KG link prediction.","feed_headline":"Query-specific graph pooling tops 42 of 48 link-prediction settings","feed_subtitle":"Selecting only logically relevant neighbors boosts RED-GNN and NBFNet on transductive and inductive benchmarks.","key_machinery":"The central object is the Context Neighbor Family (CNF) together with the two scores that define it. For a query relation $r$ and a candidate set of neighboring relations $\\mathrm{NR}'$, neighborhood precision is $\\mathrm{Rel}_{\\mathrm{pre}}(\\mathrm{NR}', r) = |\\{e : \\mathrm{NR}' \\cup \\{r\\} \\subseteq \\mathrm{NR}_e\\}| \\,/\\, |\\{e : \\mathrm{NR}' \\subseteq \\mathrm{NR}_e\\}|$, and neighborhood recall is $\\mathrm{Rel}_{\\mathrm{rec}}(\\mathrm{NR}', r) = |\\{e : \\mathrm{NR}' \\cup \\{r\\} \\subseteq \\mathrm{NR}_e\\}| \\,/\\, |\\{e : r \\in \\mathrm{NR}_e\\}|$. CNF collects all relation sets whose combined relevance score clears a threshold. Because the scores depend only on relations, they transfer to unseen entities. The optimized Algorithm 3 invokes Assumption 1, pairwise and conditional independence given the query relation, so that combined scores factor into products of per-relation scores; the chosen neighbors at each hop are the intersection of the current entity's neighboring relations with the stored high-scoring set. This per-hop selection acts as the score generator, node selector, and graph coarsening of a node-drop pooling operation specialized to heterogeneous knowledge graphs.","core_discovery":"The paper's central discovery is that the logically relevant neighbors of a query can be identified without entity embeddings or a schema, using only counts of relation co-occurrences. For a query relation $r$, the context neighbor family $\\mathrm{CNF}(r)$ is the collection of relation sets $\\mathrm{NR}'$ whose neighborhood precision and neighborhood recall exceed a threshold. Neighborhood precision is the fraction of entities whose neighborhood contains $\\mathrm{NR}'$ that also contain $r$; neighborhood recall is the fraction of entities containing $r$ that also contain $\\mathrm{NR}'$. Because these scores are computed from the training graph and depend only on relations, they transfer to unseen entities. The optimized version of the method (Algorithm 3) assumes that any two non-redundant neighboring relations are independent given the query relation; then combined scores factorize, and the best neighbor set for a query is simply the intersection of the current entity's neighboring relations with the stored high-scoring relations. The paper reports that this query-specific pooling lifts RED-GNN and NBFNet to state-of-the-art MRR and Hit@1 in 42 of 48 settings, with paired t-test $p$-values below $10^{-3}$ for transductive MRR and Hit@1 and for inductive MRR, and $2 \\times 10^{-3}$ for inductive Hit@1.","pith_inferences":["Inference: the same relation-co-occurrence statistics could be used to pre-compute a per-relation support graph before training, shrinking the effective graph for memory-bound GNNs; the paper does not test this.","Inference: because the pooling layer is unsupervised and entity-free, it should transfer to other relation-heavy graphs beyond knowledge graphs, such as heterogeneous social or biological networks with typed edges.","Inference: the ablation suggests a tunable trade-off — on small relation sets the exact power-set search is preferable (FB15k-237-V1 MRR 0.396 versus 0.383 for the optimized version), while on large relation sets only the factorized version is feasible; a hybrid that uses exact search on small relation neighborhoods could recover the lost accuracy."],"forward_implications":["RED-GNN+CP and NBFNet+CP can be dropped into existing GNN pipelines with only two thresholds to set; on 88 of 96 transductive and inductive settings the pooled version beats the base model.","Pooling cost is $O(|R_G|^2)$ for training and $O(|\\mathrm{CNF}'|)$ per query, so the scalability bottleneck remains the GNN itself.","Because the selection is relation-based, the same context neighbor family learned on the training graph applies to test entities that were never seen at training time.","On WN18RR-V2 in the transductive setting the gain over NBFNet is 11.7% in MRR and 19.4% in Hit@1; on NELL-995-V4 in the inductive setting the gain over RED-GNN is 11.7% in MRR and 16.8% in Hit@1."],"supporting_citations":[{"why":"Provides NBFNet, one of the two base GNN models that Context Pooling wraps and benchmarks against.","marker":"[33]"},{"why":"Provides RED-GNN, the second base model, and the all-entities-as-negatives evaluation configuration used in the experiments.","marker":"[31]"},{"why":"Releases the four versions of WN18RR, FB15k-237, and NELL-995 used for transductive and inductive testing, and supplies the GRAIL baseline.","marker":"[21]"},{"why":"Evidence that vanilla aggregation in GNNs does not significantly help KG link prediction, which motivates selecting neighbors.","marker":"[8]"},{"why":"Shows that replacing aggregation with an MLP yields comparable results, the negative result Context Pooling is designed to overcome.","marker":"[32]"},{"why":"Markov-chain reference invoked to support the conditional-independence factorization in Theorem 1.","marker":"[16]"},{"why":"Supplies the graph-pooling taxonomy (flat versus hierarchical, node-drop) that Context Pooling positions itself within.","marker":"[12]"}],"fun_headline_variants":["Pooling by query relevance wins 42 of 48 link-prediction tasks","Context Pooling: query-specific neighbor selection hits SOTA in 42/48","Relation-only pooling lifts GNN link prediction to SOTA in 42/48","Query-specific graph pooling: 42/48 SOTA results in KG link prediction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"All reported results use Algorithm 3, whose correctness relies on Assumption 1: that any two non-redundant neighboring relations occur independently of each other, and independently given the query relation; the paper's own ablation shows this assumption is approximate, since exact search reaches higher MRR on FB15k-237-V1.","fun_headline_variants_meta":{"raw":{"variants":["Pooling by query relevance wins 42 of 48 link-prediction tasks","Context Pooling: query-specific neighbor selection hits SOTA in 42/48","Relation-only pooling lifts GNN link prediction to SOTA in 42/48","Query-specific graph pooling: 42/48 SOTA results in KG link prediction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00023,"raw_usage":{"total_tokens":1509,"prompt_tokens":1001,"completion_tokens":508,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":617,"completion_tokens_details":{"reasoning_tokens":421}},"tokens_in":617,"tokens_out":508,"duration_ms":5292,"temperature":1.0,"reasoning_tokens":421,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:37:36.445149+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct or select a KG where the co-occurrence of neighboring relations is strongly correlated with the query relation, for instance a domain where the relation Profession usually co-occurs with Award only when Company is also present, and run RED-GNN+CP with Algorithm 3 versus plain RED-GNN. If the pooled model's MRR drops below the base model while the unoptimized Algorithm 1 does not drop, the independence assumption is the load-bearing approximation.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides NBFNet, one of the two base GNN models that Context Pooling wraps and benchmarks against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides RED-GNN, the second base model, and the all-entities-as-negatives evaluation configuration used in the experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Releases the four versions of WN18RR, FB15k-237, and NELL-995 used for transductive and inductive testing, and supplies the GRAIL baseline."},{"cited_title":"Are Message Passing Neural Networks Really Helpful for Knowledge Graph Completion?","cited_arxiv_id":"2205.10652","evidence_quote":"Evidence that vanilla aggregation in GNNs does not significantly help KG link prediction, which motivates selecting neighbors."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Shows that replacing aggregation with an MLP yields comparable results, the negative result Context Pooling is designed to overcome."},{"cited_title":"1998.Markov chains","cited_arxiv_id":null,"evidence_quote":"Markov-chain reference invoked to support the conditional-independence factorization in Theorem 1."}],"review_version":1}