{"id":"692af1b2-4e73-4ee9-bd99-fca0cdc7e48f","arxiv_id":"2412.16250","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"FreeHGC performs training-free heterogeneous graph condensation through structural data selection and synthesis, matching or beating training-based condensation on seven datasets.","lead":"FreeHGC shrinks large heterogeneous graphs by picking the most informative nodes and blending the rest, without training a model during condensation. It reports near-full-graph accuracy using less than 5% of the original nodes, and is much faster than prior training-based condensation methods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Root/father/leaf type decomposition is asserted but never specified for complex datasets, leaving Freebase/AM results unreproducible and the general claim unestablished.","rationale":"The reader's weakest_assumption identifies exactly this root/father/leaf decomposition as unchecked for complex topologies. I agree that this is the most load-bearing concern: it directly affects the validity of the method on four of the seven reported datasets and the reproducibility of the central claim. The submodularity and Algorithm 1 issues noted by the reader are real but secondary—they concern the theoretical wrapper, not the empirical core, and could be fixed by rephrasing without changing the algorithm. The father/leaf assumption is part of the algorithm itself and is under-specified in the text. I credit the paper for releasing code, reporting extensive experiments, and providing ablations on three datasets; those are strong positives. But the missing specification and missing ablation on complex datasets mean the current CONDITIONAL verdict remains appropriate. The concrete test I propose would resolve whether the concern actually lands: if the code's role assignment is principled and the ablation shows both components are necessary on Freebase/AM, the concern is answered; if not, the paper should be revised to specify the assignment and justify it empirically.","tokens_in":23728,"tokens_out":6700,"duration_ms":59987,"concrete_test":"Inspect the released code to recover the exact father/leaf role assignment for Freebase, AM, and MUTAG, then run the Table VIII ablation variants (Variant#4, #5, #6) on those datasets. If removing either the father-type PageRank selection or the leaf-type mean-aggregation synthesis changes accuracy by more than roughly two points, or if an alternative role assignment (e.g., swapping father/leaf roles) produces materially different results, the root/father/leaf decomposition is a hidden, load-bearing design choice rather than a validated structural assumption.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The second component of FreeHGC (Section IV-C, Algorithm 2) rests on a topological premise: every heterogeneous graph can be partitioned into root, father, and leaf types, where father types bridge root and leaf. The paper states this classification in prose and Figure 5, but it never defines how father vs. leaf roles are assigned for datasets with many types, such as Freebase (8 types), AM (7 types), and MUTAG (7 types). The ablation study in Table VIII covers only ACM, DBLP, and AMiner, so the necessity and correctness of the father/leaf split is empirically unchecked exactly on the datasets where the assumption is least obvious. If the role assignment is ad hoc or manually curated per dataset, then the reported gains on Freebase and AM are not reproducible from the paper description, and the claimed general training-free heterogeneous graph condensation framework is not established. This concern is load-bearing because the two proposed other-type condensation strategies—Personalized PageRank selection for father types and mean-aggregation synthesis for leaf types—are the mechanism by which the method claims to preserve graph structure information; an unspecified or arbitrary decomposition undermines that mechanism.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FreeHGC, a training-free method for heterogeneous graph condensation. Target-type nodes are selected with a unified score combining receptive-field coverage and a meta-path Jaccard diversity term; other-type nodes are condensed with Personalized PageRank for 'father' types and mean-aggregation synthesis for 'leaf' types. Experiments on ACM, DBLP, IMDB, Freebase, AMiner, MUTAG, and AM report higher accuracy than HGCond and coreset baselines at most condensation ratios, better generalization across HGB/HGT/HAN/SeHGNN, and large speedups. The paper also claims a submodularity-based approximation guarantee for the selection criterion.","tokens_in":23970,"tokens_out":5460,"duration_ms":46379,"significance":"If the results hold, FreeHGC is a useful contribution: it decouples condensation from trained relay models and makes condensation much faster and model-agnostic. The empirical evaluation is broad (seven datasets, including one large-scale), reports mean plus variance, includes generalization to four HGNN models, and the code is public, which are genuine strengths. The core empirical claim is credible. However, two load-bearing gaps—an unspecified father/leaf decomposition for complex datasets and a mismatch between the stated submodularity theory and the implemented algorithm—must be addressed before the paper's full claims are acceptable.","major_comments":[{"comment":"The theoretical claim is not matched by the implementation. The paper states that F(S) is submodular and that greedy maximization gives a (1−1/e) approximation, but Algorithm 1 (line 10) does not perform greedy set maximization; it computes per-node scores and selects the top-k. The greedy guarantee therefore does not apply to the actual algorithm. In addition, submodularity of 1−J(S) is asserted for the specific aggregate Jaccard measure of Eqs. (5)–(7) rather than proved. This does not invalidate the empirical results, but the theoretical section must either be aligned with the implementation (e.g., by presenting a genuinely greedy variant) or the (1−1/e) approximation claim should be removed or explicitly made conditional on a different selection procedure.","section":"§IV-B, Eq. (8) and Algorithm 1"},{"comment":"The root/father/leaf decomposition is not specified for datasets with many node types, which is load-bearing because the two other-type condensation strategies are defined only after this decomposition. The paper never states which node types of Freebase (8 types), MUTAG (7 types), or AM (7 types) are treated as father types versus leaf types, nor does it give a general rule for assigning these roles. The ablation study in Table VIII covers only ACM, DBLP, and AMiner, so the necessity and correctness of the father/leaf split are empirically unchecked exactly on the datasets where the topology is least obvious. As written, the Freebase, MUTAG, and AM results are not reproducible from the paper description; the authors should provide the per-dataset type mappings or a precise algorithm that derives them automatically.","section":"§IV-C, Figure 5, and Algorithm 2"},{"comment":"The stated complexity for condensing target-type nodes, O(α N_tgt^2 + N_tgt log N_tgt), is not derived and is not credible for the reported large-scale experiments. For AMiner, N_tgt ≈ 4.89×10^6 and α ranges from 0.0005 to 0.008, which gives roughly 10^10 to 10^11 operations under this formula; this is inconsistent with the modest condensation times shown in Figure 8. The authors should either identify which step causes the quadratic term, provide a sparse/approximate implementation that avoids it, or correct the complexity expression.","section":"§IV-C, Time Complexity"}],"minor_comments":[{"comment":"The reported average for Herding-HG on Freebase is 40.29, but the four preceding accuracies (49.81, 47.22, 43.44, 51.17) average to approximately 47.91; please correct this arithmetic or transcription error.","section":"Table IV, Freebase row"},{"comment":"The caption says 'Left: ACM dataset. Left: IMDB dataset.'; the second label should be 'Right'.","section":"Figure 7 caption"},{"comment":"The text repeatedly uses 'corset methods'; the standard term is 'coreset methods'.","section":"§V-A"},{"comment":"The text says 'Figure VI shows that FreeHGC performs best', but the referenced results are in Table VI; please correct the cross-reference.","section":"§V-G"}],"recommendation":"major_revision","confidential_remarks":"The code release may already contain the missing per-dataset type mappings and the actual selection procedure; if so, the authors should document these in the paper rather than leaving them implicit. I would not recommend rejection because the empirical contribution appears solid and the issues seem addressable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is the first training-free method for heterogeneous graph condensation and the empirical core is credible. It deserves a serious referee, but the theoretical wrapper and the topology-typing step need real work.\n\nWhat's new: FreeHGC reframes heterogeneous condensation as data selection. For target-type nodes it combines receptive-field coverage with a Jaccard diversity term across generated meta-paths; for other types it uses Personalized PageRank to pick important 'father' nodes and mean-aggregated synthesis for 'leaf' nodes. That combination is genuinely new in the cited literature, and the paper shows it works. The results are extensive: accuracy and variance across seven datasets, generalization across four HGNNs, a large-scale AMiner run at 0.8% condensation reaching 91.86% versus 87.02% for HGCond, and big speedups in condensation time. The ablation study is reasonable, and the code is promised. The empirical core holds up.\n\nSoft spots, in proportion: the theory section is not what the algorithm does. Algorithm 1 takes top-k scores per meta-path, not the greedy submodular maximization the text claims, so the (1-1/e) guarantee doesn't apply to the implemented method. The claim that 1-J(S) is submodular is asserted via references rather than shown for the paper's node-set definition. PPR's alpha is never reported. None of these kill the accuracy results, but they need fixing.\n\nThe bigger issue is the root/father/leaf decomposition. For ACM/DBLP/AMiner the roles are intuitive, and the ablation shows the two other-type strategies matter. But Freebase has 8 node types and AM has 7, and the paper never specifies how a type is assigned to father versus leaf. The ablation doesn't cover those datasets. So the framework's generality—the claim that it works on complex topologies—is not established, and reproducing Freebase/AM results from the description would be hard. This is load-bearing for the 'general framework' claim, though it doesn't undermine the reported numbers.\n\nBottom line: this is a useful, solidly evaluated method with an overreaching theoretical narrative. Send it to review, ask for the greedy/top-k discrepancy, the PPR details, and an explicit type-role assignment rule. I'd cite it as a baseline if I worked on graph condensation.","headline":"First training-free heterogeneous graph condensation with a strong empirical showing; the theory section and the topology decomposition need cleanup before I'd trust the framework's generality.","tokens_in":24472,"tokens_out":2192,"would_cite":true,"duration_ms":19724,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","05C85"],"pacs":[],"model":"deepseek-v4-flash","headline":"FreeHGC claims that heterogeneous graph condensation can be done without any model training, and that this training-free approach beats the trained baseline HGCond in accuracy, speed, and generalization.","keywords":["heterogeneous graph condensation","training-free graph condensation","data selection","heterogeneous graph neural networks","meta-path","submodular optimization","Personalized PageRank","node classification"],"falsifier":"Run FreeHGC on a heterogeneous graph where the father and leaf roles can be permuted among the non-target types; if the best role assignment varies unpredictably with the dataset and FreeHGC's advantage over random sampling disappears under another assignment, the root-father-leaf premise is falsified.","tokens_in":23545,"feed_emoji":"🕸️","tokens_out":5615,"duration_ms":43870,"temperature":0.7,"pith_summary":"This paper argues that heterogeneous graph condensation does not need to train a relay model: it can be reframed as a data selection problem. FreeHGC scores target-type nodes by combining receptive-field coverage with meta-path diversity, then condenses the remaining node types with Personalized PageRank and mean-aggregation synthesis. The claim is that this training-free procedure beats the existing trained method HGCond in accuracy, speed, and generalization across seven heterogeneous graph datasets, while allowing the condensation ratio to grow without the overfitting degradation seen in trained condensation. If true, graph condensation becomes a fast pre-processing step that any heterogeneous graph neural network can consume.","feed_headline":"No-training graph condensation wins on seven benchmarks","feed_subtitle":"Data selection with receptive fields and meta-path diversity delivers higher accuracy in a fraction of the time.","key_machinery":"The load-bearing object is the unified data selection score for target nodes, $F(S) = R(S)/|\\hat{R}| + (1 - J(S))$, where $R(S)$ is the size of the union of receptive fields activated by the selected nodes under a given meta-path and $J(S)$ is the normalized Jaccard similarity of that meta-path with all other meta-paths. Because both summands are argued to be submodular, the score enables a greedy selection with an approximation guarantee. The second piece is the root-father-leaf taxonomy of node types: father types are condensed by ranking nodes with Personalized PageRank on symmetrized meta-path adjacency matrices, and leaf types are synthesized by mean-aggregating the features of each father's neighbors, with reverse edges added to preserve father-to-father two-hop information.","core_discovery":"FreeHGC's central claim is that a training-free, data-selection approach can outperform the trained gradient-matching method HGCond for heterogeneous graph condensation. The target-type selection score $F(S) = R(S)/|\\hat{R}| + (1 - J(S))$ combines submodular receptive-field maximization with a Jaccard-based meta-path diversity term, and the greedy algorithm inherits the $(1 - 1/e)$ approximation guarantee for submodular maximization. Other-type nodes are condensed by treating them as 'father' nodes selected by Personalized PageRank importance and 'leaf' nodes synthesized by mean aggregation, based on the observation that neighbor attention in HGNNs is dispensable. Across ACM, DBLP, IMDB, Freebase, MUTAG, AM, and AMiner, FreeHGC reports higher test accuracy than HGCond and coreset baselines, at a fraction of the condensation time and with accuracy that keeps rising as the condensation ratio grows.","pith_inferences":["This suggests that heterogeneous graph condensation with a fixed budget is an influence-maximization problem on a multiplex of meta-path adjacency matrices, which may admit coreset-style guarantees beyond node classification.","The root-father-leaf taxonomy is a hidden prior; for knowledge graphs with eight or more node types and many relations, defining father and leaf roles may require per-dataset design, and automatic role discovery would be a natural extension.","The mean-aggregation synthesis step ties the method to the finding that neighbor attention is unnecessary; if a future HGNN relies on fine-grained neighbor weighting, FreeHGC's condensed leaf nodes would likely need to store richer statistics than a mean.","PageRank-based father selection could be replaced or augmented by other importance measures, so the framework is a template rather than a single fixed algorithm."],"forward_implications":["Heterogeneous graph condensation becomes a pre-processing routine: no relay model, no bi-level optimization, no nested training loops, and no clustering or OPS initialization.","Accuracy improves as the condensation ratio grows, so practitioners can pick any budget without the overfitting-induced performance drops reported for HGCond.","The condensed graph transfers across meta-path-based and meta-path-free HGNNs (HAN, HGT, HGB, SeHGNN), so one condensed graph can serve many downstream models.","On the tested graphs the procedure cuts storage by roughly 95 to 99 percent and reduces downstream training time to a small fraction of whole-graph training."],"supporting_citations":[{"why":"HGCond is the trained heterogeneous graph condensation baseline whose limitations in effectiveness, efficiency, and generalization motivate FreeHGC.","marker":"[26]"},{"why":"GCond supplies the gradient-matching paradigm and bi-level optimization that HGCond follows, and which FreeHGC avoids entirely.","marker":"[24]"},{"why":"SeHGNN provides the empirical finding that neighbor attention is dispensable (mean aggregation suffices) and the general meta-path generation scheme that FreeHGC adapts.","marker":"[18]"},{"why":"Nemhauser et al. provides the greedy approximation guarantee for submodular maximization that justifies the target-node selection criterion.","marker":"[50]"},{"why":"Approximate Personalized PageRank supplies the scalable importance computation used to condense father-type nodes.","marker":"[56]"},{"why":"The HGB benchmark provides four of the datasets (ACM, DBLP, IMDB, Freebase) and the evaluation protocol used in the experiments.","marker":"[8]"}],"fun_headline_variants":["Data selection beats training for graph condensation","Training-free method condenses heterogeneous graphs faster","FreeHGC: no-training graph condensation with higher accuracy","Data selection outperforms trained graph condensation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that every heterogeneous graph can be cleanly split into root, father, and leaf node types, and that condensing fathers by Personalized PageRank plus synthesizing leaves by mean aggregation preserves the structure that matters.","fun_headline_variants_meta":{"raw":{"variants":["Data selection beats training for graph condensation","Training-free method condenses heterogeneous graphs faster","FreeHGC: no-training graph condensation with higher accuracy","Data selection outperforms trained graph condensation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000112,"raw_usage":{"total_tokens":1084,"prompt_tokens":995,"completion_tokens":89,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":611,"completion_tokens_details":{"reasoning_tokens":45}},"tokens_in":611,"tokens_out":89,"duration_ms":1836,"temperature":1.0,"reasoning_tokens":45,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T11:21:53.970810+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run FreeHGC on a heterogeneous graph where the father and leaf roles can be permuted among the non-target types; if the best role assignment varies unpredictably with the dataset and FreeHGC's advantage over random sampling disappears under another assignment, the root-father-leaf premise is falsified.","supporting_citations":[{"cited_title":"Heterogeneous graph condensation,","cited_arxiv_id":null,"evidence_quote":"HGCond is the trained heterogeneous graph condensation baseline whose limitations in effectiveness, efficiency, and generalization motivate FreeHGC."},{"cited_title":"Graph condensation for graph neural networks,","cited_arxiv_id":null,"evidence_quote":"GCond supplies the gradient-matching paradigm and bi-level optimization that HGCond follows, and which FreeHGC avoids entirely."},{"cited_title":"Simple and efficient het- erogeneous graph neural network,","cited_arxiv_id":null,"evidence_quote":"SeHGNN provides the empirical finding that neighbor attention is dispensable (mean aggregation suffices) and the general meta-path generation scheme that FreeHGC adapts."},{"cited_title":"An analysis of approximations for maximizing submodular set functions - I,","cited_arxiv_id":null,"evidence_quote":"Nemhauser et al. provides the greedy approximation guarantee for submodular maximization that justifies the target-node selection criterion."},{"cited_title":"Scaling graph neural networks with approximate pagerank,","cited_arxiv_id":null,"evidence_quote":"Approximate Personalized PageRank supplies the scalable importance computation used to condense father-type nodes."},{"cited_title":"Are we really making much progress?: Revisiting, benchmarking and refining heterogeneous graph neural networks,","cited_arxiv_id":null,"evidence_quote":"The HGB benchmark provides four of the datasets (ACM, DBLP, IMDB, Freebase) and the evaluation protocol used in the experiments."}],"review_version":1}