{"id":"6c5ebb48-573c-41f5-8f6d-c00029b5006e","arxiv_id":"2505.08157","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A knowledge-aware recommender combining Lorentz-model hyperbolic embeddings with three model-level contrastive augmentations reports up to 11.03% improvement over the strongest baseline.","lead":"This paper builds a recommender that learns user and item embeddings in hyperbolic space and creates contrastive training views by altering the model itself rather than the interaction graph. The best variant reports up to 11.03% improvement over strong baselines on three public datasets.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Single-run comparisons with no significance testing leave the 'always ranks first' claim unsupported; multi-seed reruns are needed.","rationale":"The reader's CONDITIONAL verdict is reasonable, and the reader correctly flags missing error bars as part of the rationale. However, the reader's designated weakest assumption is the unmeasured preference-shift premise of the model augmentations. I regard the statistical reliability of Table 2 as the more load-bearing issue for the central claim: the paper's headline is an empirical superiority claim, and several reported margins are below 2%, a range where seed variance alone can reverse rankings in contrastive GNN recommenders. A multi-seed significance rerun directly tests the central claim without requiring new theory. If it fails, the paper's main conclusion cannot stand; if it passes, the remaining weaknesses (unmeasured preference shift, missing hyperbolic baselines, citation hygiene) are addressable in revision. Therefore the reader's CONDITIONAL verdict should remain unchanged rather than being hardened or relaxed.","tokens_in":14453,"tokens_out":8656,"duration_ms":81418,"concrete_test":"Run at least 5 independent seeds of HCMKR-C, HCMKR-D, HCMKR-P and the strongest baselines (KGCL, SGL, LightGCN) under the exact Table 2 protocol, reporting mean and standard deviation for every metric on all three datasets, and apply paired significance tests between HCMKR-C and the per-metric best baseline. If any claimed improvement below roughly 2% (e.g., Yelp R@20 0.26% for HCMKR-P, Amazon N@10 1.38%) is not significant at p<0.05, the 'always ranks first' statement in Section 4.2 is not supported as written.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.2 stakes the central claim on Table 2: HCMKR-C 'always ranks first' with gains of 2.33% to 11.03% over the per-metric best baseline. The experiments section reports no number of seeds, no standard deviations, and no significance tests anywhere in the protocol. On Yelp2018, HCMKR-C's smallest margin is 2.33% (N@10) and HCMKR-P's R@20 margin is only 0.26%; on Amazon-Book, HCMKR-P's N@10 margin is 1.38% and R@10 margin is 0.49% over SGL. These magnitudes are within the seed-to-seed variance commonly observed in contrastive GNN recommenders. Since 'always ranks first' is the central claim, the absence of uncertainty quantification is the most load-bearing gap. The preference-shift argument in Section 3.4 is asserted rather than measured, and the omission of hyperbolic baselines (e.g., HGCF) weakens external anchoring, but both are secondary: if the Table 2 margins are not reproducible, the headline empirical claim collapses regardless of whether the mechanism story is true.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes HCMKR, a knowledge-aware recommender that represents users and items in Lorentzian hyperbolic space and applies three model-level augmentations (dropout, cross-layer outputs, magnitude pruning) to generate positive pairs for contrastive learning. The central claim is that, unlike structure-level augmentations such as edge dropping, these model-level augmentations avoid preference shift, and that the hyperbolic aggregation captures the hierarchical structure of user-item and knowledge graphs. The paper reports experiments on Yelp2018, Amazon-Book, and MovieLens-20M, claiming that HCMKR-C always ranks first with gains of 2.33% to 11.03% over the best baseline, supported by ablations, parameter-sensitivity plots, and a runtime comparison.","tokens_in":14769,"tokens_out":4342,"duration_ms":39257,"significance":"If the empirical claims hold, the paper makes two useful contributions: it introduces model-level augmentation as a principled alternative to structural augmentation in knowledge-aware contrastive recommendation, and it demonstrates that Lorentzian aggregation can be integrated with standard contrastive objectives without prohibitive computational cost. The manuscript provides code, the Lorentzian mathematics (Eqs. 1-12) is standard and internally consistent, and the ablation study is well structured. However, the central empirical claim is currently under-supported: the experimental protocol reports no variance estimates, and the hyperbolic component is not benchmarked against existing hyperbolic recommenders.","major_comments":[{"comment":"The headline claim that HCMKR-C 'always ranks first' with gains of 2.33% to 11.03% is not supported by the reported protocol, which contains no number of seeds, no standard deviations, and no significance tests. Margins as small as 2.33% (NDCG@10 on Yelp2018) and 1.38% (NDCG@10 on Amazon-Book for HCMKR-P) are within the seed-to-seed variation typically observed in contrastive GNN recommenders, so multi-seed runs with variance and significance testing are required.","section":"Section 4.2, Table 2"},{"comment":"The central mechanism claim that model-level augmentations avoid preference shift is asserted and illustrated but never directly measured. The paper should include a controlled comparison between model-level and structure-level augmentations under the same hyperbolic encoder, or a quantitative proxy for preference semantics (e.g., the change in top-ranked items per user before and after augmentation), to substantiate the claimed advantage over KGCL-style structural augmentation.","section":"Section 3.4, Fig. 2(e)"},{"comment":"No hyperbolic baseline is included in the comparison; existing hyperbolic recommenders such as HGCF [23] and the Poincaré-based knowledge-aware method [17] are cited in related work but absent from Table 2. Without such baselines, the contribution of the Lorentzian aggregation is not externally anchored, and the reported gains could in principle be attributed to the contrastive regularization rather than to hyperbolic geometry.","section":"Section 4, Table 2"},{"comment":"There is an internal contradiction in the pruning augmentation: the text states 'set the elements in Θ greater than the ξ to 0', but the mask M_P^2 = 1(|Θ|>ξ) keeps elements greater than ξ and zeros smaller elements. Please correct the threshold direction and clarify the pruning-ratio definition, since this determines the actual behavior of HCMKR-P and is needed for reproducibility.","section":"Section 3.4, Eq. (16)"}],"minor_comments":[{"comment":"The logarithmic map in Eq. (3) appears garbled in the typesetting; please restate the normalized vector expression clearly.","section":"Section 3.2, Eq. (3)"},{"comment":"The text describes z_i^(k) as 'the representation of the item node u'; this should be item node i.","section":"Section 3.3, after Eq. (12)"},{"comment":"The HCMKR-P Improve row is N/A on MovieLens-20M even though several metrics improve; report the comparison against KGCL explicitly, and note that R@20 is actually lower than KGCL on that dataset.","section":"Table 2"},{"comment":"The claim of 'reduced time consumption (×0.88)' applies only to HCMKR-C; Table 3 shows HCMKR-P and HCMKR-D are slower than KGCL (×1.28 and ×1.27). Please qualify the efficiency claim accordingly.","section":"Contributions bullet and Table 3"},{"comment":"The denominator of the contrastive loss uses the set j∈{i+}∪V\\{i}; please clarify whether this includes the positive sample itself and whether duplicate indices are excluded, to make the loss unambiguous.","section":"Section 3.5, Eq. (19)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of the journal, and the core ideas are publishable if the experimental rigor is improved. The main barrier is the absence of uncertainty quantification and significance testing in Table 2, together with the missing hyperbolic baselines and the untested preference-shift mechanism. I recommend asking the authors to provide multi-seed experiments with error bars, add at least one hyperbolic baseline, add a direct test of the preference-shift claim, and fix the pruning-mask inconsistency before resubmission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this is a competent integration of existing ideas—hyperbolic (Lorentz) GNNs for knowledge-aware recommendation plus three model-level contrastive augmentations. It reports large gains on three standard benchmarks, and the math and tables are internally consistent. The soft spot is not the method but the evidence: single-run numbers, no error bars, no hyperbolic baselines. The 'always ranks first' claim is plausible but not yet airtight.\n\nWhat's new: the specific combination of relation-aware Lorentzian knowledge aggregation and model-level augmentations (dropout, cross-layer outputs, magnitude pruning) to avoid preference shift. The components exist separately (HGCF, SimGRACE, XSimGCL, KGCL), but the paper's contribution is the integration, plus the cross-layer view in hyperbolic space. The code is available, which is a plus. The ablation (Fig. 5) shows both hyperbolic geometry and model-augmentation contribute, and the parameter sensitivity analysis is reasonable.\n\nSoft spots, in order of importance. First, no seed-level variance or significance tests anywhere. The smallest margins in Table 2 are 0.26% (Yelp R@20 for HCMKR-P) and 0.49% (Amazon-Book R@10 for HCMKR-P); those are within the seed-to-seed noise I'd expect from contrastive GNN recommenders. HCMKR-C's gains are larger, but the paper treats all numbers as deterministic. Multi-seed reruns would fix this and should be the first request. Second, the paper argues hyperbolic geometry captures hierarchy but never compares against a hyperbolic recommender (HGCF and the fully hyperbolic GCN are cited but not run as baselines). That leaves the mechanism story under-supported. Third, the preference-shift argument in Section 3.4 is asserted from a cartoon rather than measured. It is plausible, and the results are consistent with it, but it isn't direct evidence.\n\nThat said, the architecture is reasonable and the paper is readable. A few self-citations (Refs. 11–13) look tangential, but nothing that changes the result.\n\nWho this is for: people working on knowledge-aware recommendation or contrastive learning for GNNs. It deserves a serious referee; the empirical claim, if reproducible, is a new SOTA on three benchmarks. My recommendation: send to review, with a request for multi-seed results and at least one hyperbolic baseline.","headline":"A sensible integration paper whose SOTA claim needs multi-seed evidence and a hyperbolic baseline before the margins can be trusted.","tokens_in":15221,"tokens_out":2726,"would_cite":true,"duration_ms":27795,"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":"Replacing structural graph perturbation with model-level augmentation—dropout, cross-layer outputs, and pruning—and embedding in Lorentzian hyperbolic space improves knowledge-aware recommendation by up to 11.03%.","keywords":["knowledge-aware recommendation","hyperbolic representation learning","Lorentz model","contrastive learning","model augmentation","knowledge graph","preference shift","graph neural networks"],"falsifier":"Measure preference drift directly: for a fixed user, compare the ranking of held-out items induced by the unaugmented representation with the ranking induced by each augmented view. If the edge-dropped view shifts the ranking toward irrelevant items no more than the model-augmented views do, the preference-shift premise fails.","tokens_in":14266,"feed_emoji":"📈","tokens_out":9125,"duration_ms":82651,"temperature":0.7,"pith_summary":"The paper proposes HCMKR, a knowledge-aware recommender that learns user and item embeddings in the Lorentz model of hyperbolic space and builds contrastive positive pairs by perturbing the model rather than the graph. It aims to establish two things: hyperbolic aggregation captures the hierarchical, power-law structure of real interaction and knowledge graphs better than Euclidean encoders, and model-level augmentations—dropout masks, cross-layer outputs, and magnitude pruning—create positive views without shifting user preferences the way edge dropping does. On Yelp2018, Amazon-Book, and MovieLens-20M, the cross-layer variant HCMKR-C reports gains of 2.33% to 11.03% over the best baseline, with HCMKR-D close behind. If these results hold, structural graph augmentation is unnecessary for contrastive recommendation, and hyperbolic geometry plus model perturbation is a cheaper route to the same self-supervision.","feed_headline":"Perturb the model, not the graph: recommendation up to 11% better","feed_subtitle":"Hyperbolic contrastive learning builds positive pairs from dropout, layer outputs, and pruning instead of deleting edges.","key_machinery":"The load-bearing machinery is the Lorentz model of hyperbolic geometry, used through exponential and logarithmic maps so that aggregation happens in the tangent space while embeddings live on the manifold of negative curvature; this is supported by a relation-aware attention weight based on a Fermi-Dirac function of hyperbolic distance. The second piece is the family of model-level augmentations: dropout masks (two forward passes of the same input), cross-layer outputs (views from different GNN layers in a single pass), and magnitude pruning (zeroing weights under a threshold). Together they replace structural edge dropping, which the paper argues rewrites a user's purchase history and shifts preference semantics, with perturbations that keep the input graph and its supervision signal unchanged.","core_discovery":"On the paper's own terms, the central discovery is that the two standard limitations of contrastive knowledge-aware recommendation can be addressed by the same design move: do the representation learning in the Lorentzian hyperbolic manifold, and generate contrastive views by altering the encoder rather than the input graph. The proposed Lorentzian knowledge aggregation propagates relation-weighted messages in the tangent space, then the hyperbolic encoder performs symmetric neighbor aggregation; the three model-level augmentations produce two views of every user and item without touching the user-item edges or knowledge-graph triplets. The reported experiments claim that HCMKR-C always ranks first across all metrics on all three datasets, with improvements from 2.33% to 11.03% over the strongest baseline, and that the learned embeddings are more uniformly distributed. The paper attributes the gain specifically to hierarchy capture from hyperbolic geometry and preference preservation from model-level augmentation.","pith_inferences":["If the preference-shift mechanism is real, the same model-augmentation recipe should transfer to graph contrastive tasks beyond recommendation, such as social or citation networks, where edge deletion also destroys node semantics.","A direct measurement of user-embedding drift under each augmentation would turn the Figure 2(e) preference-shift story into a quantitative claim.","Because the three augmentations differ in cost and effect, combining or adaptively selecting them per batch may outperform any single variant.","The per-epoch time comparison suggests efficiency, but an end-to-end comparison including tuning would be the more demanding test."],"forward_implications":["Structural edge dropping is not necessary for contrastive recommendation; model-level perturbations alone can supply the positive views.","A knowledge-aware recommender can run in hyperbolic space without a prohibitive time penalty: HCMKR-C trains at about 0.88 times the per-epoch cost of KGCL on the tested datasets.","Taking different GNN layer outputs as the two contrastive views is the strongest of the three augmentations, suggesting that depth itself provides a useful view separation.","The gains are largest on MovieLens-20M, the largest dataset here, so the benefit of hyperbolic modeling and model augmentation appears to scale with interaction volume."],"supporting_citations":[{"why":"Supplies the main baseline (KGCL) whose structural knowledge-graph augmentations the paper replaces with model-level augmentations.","marker":"[35]"},{"why":"Defines the knowledge-aware propagation and dataset split that the proposed aggregation extends to hyperbolic space.","marker":"[30]"},{"why":"Shows that feature transformation and nonlinear activation can be dropped in recommendation GNNs, which the hyperbolic encoder follows.","marker":"[7]"},{"why":"Introduces the Lorentz model and its distance, chosen here for numerical stability over the Poincaré ball.","marker":"[14]"},{"why":"Provides the exponential/logarithmic space-mapping recipe and trainable curvature used for hyperbolic embeddings.","marker":"[2]"},{"why":"Provides the tangent-space aggregation pattern for hyperbolic collaborative filtering that the paper adopts.","marker":"[23]"},{"why":"SGL is the contrastive baseline whose structural augmentations (edge dropping) the paper argues cause preference shift.","marker":"[32]"},{"why":"MCCLK is the multi-level contrastive baseline whose time cost is compared against.","marker":"[38]"},{"why":"Bourgain-type embedding results are cited to argue Euclidean space poorly represents power-law graph structure.","marker":"[16]"}],"fun_headline_variants":["Model views, not edge drops: hyperbolic contrastive learning lifts rec by 11%","Hyperbolic space + model augmentation: knowledge-aware rec up to 11% better","No graph perturbation: hyperbolic contrastive learning uses model views for 11% gain","Drop edges? No: hyperbolic contrastive learning augments the model, not the graph","Hyperbolic contrastive learning: perturb the encoder, not the graph, for 11% gain"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The key assumption is that randomly dropping neurons, using different layers' outputs, and pruning small weights create alternative versions of a user's representation that still reflect the same interests; the paper does not measure that directly.","fun_headline_variants_meta":{"raw":{"variants":["Model views, not edge drops: hyperbolic contrastive learning lifts rec by 11%","Hyperbolic space + model augmentation: knowledge-aware rec up to 11% better","No graph perturbation: hyperbolic contrastive learning uses model views for 11% gain","Drop edges? No: hyperbolic contrastive learning augments the model, not the graph","Hyperbolic contrastive learning: perturb the encoder, not the graph, for 11% gain"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001419,"raw_usage":{"total_tokens":5710,"prompt_tokens":911,"completion_tokens":4799,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":527,"completion_tokens_details":{"reasoning_tokens":4688}},"tokens_in":527,"tokens_out":4799,"duration_ms":28080,"temperature":1.0,"reasoning_tokens":4688,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:02:35.068739+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure preference drift directly: for a fixed user, compare the ranking of held-out items induced by the unaugmented representation with the ranking induced by each augmented view. If the edge-dropped view shifts the ranking toward irrelevant items no more than the model-augmented views do, the preference-shift premise fails.","supporting_citations":[{"cited_title":"In: Proc","cited_arxiv_id":null,"evidence_quote":"Supplies the main baseline (KGCL) whose structural knowledge-graph augmentations the paper replaces with model-level augmentations."},{"cited_title":"In: Proc","cited_arxiv_id":null,"evidence_quote":"Defines the knowledge-aware propagation and dataset split that the proposed aggregation extends to hyperbolic space."},{"cited_title":"In: Proc","cited_arxiv_id":null,"evidence_quote":"Shows that feature transformation and nonlinear activation can be dropped in recommendation GNNs, which the hyperbolic encoder follows."},{"cited_title":"In: Proc","cited_arxiv_id":null,"evidence_quote":"Introduces the Lorentz model and its distance, chosen here for numerical stability over the Poincaré ball."},{"cited_title":"In: Advances in Neural Information Processing Systems (2019)","cited_arxiv_id":null,"evidence_quote":"Provides the exponential/logarithmic space-mapping recipe and trainable curvature used for hyperbolic embeddings."},{"cited_title":"In: Proc","cited_arxiv_id":null,"evidence_quote":"Provides the tangent-space aggregation pattern for hyperbolic collaborative filtering that the paper adopts."},{"cited_title":"In: Proc","cited_arxiv_id":null,"evidence_quote":"SGL is the contrastive baseline whose structural augmentations (edge dropping) the paper argues cause preference shift."},{"cited_title":"In: Proc","cited_arxiv_id":null,"evidence_quote":"MCCLK is the multi-level contrastive baseline whose time cost is compared against."},{"cited_title":"Combinatorica (1995)","cited_arxiv_id":null,"evidence_quote":"Bourgain-type embedding results are cited to argue Euclidean space poorly represents power-law graph structure."}],"review_version":1}