{"id":"d6b4d4d7-5a52-41ec-9d3c-82e2f439ca67","arxiv_id":"2504.13983","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"QuatE-D replaces the inner-product scoring of QuatE with Euclidean distance after a quaternion Hamilton-product rotation, and reports improved Mean Rank on WN18, FB15k, WN18RR, and FB15k-237.","lead":"QuatE-D is a knowledge graph embedding model that scores triples by the Euclidean distance between a rotated head entity and the tail entity in quaternion space. It reports competitive link prediction numbers on four standard benchmarks, with the largest gains in Mean Rank.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline MR results come from 'QuatE-D2 with type constraints' that are never defined; without knowing whether they are a test-time filter, the central empirical claim is uninterpretable.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: QuatE-D2, the variant responsible for the headline MR improvements, is defined only by the phrase 'with type constraints' in the table caption. No definition appears in the body, and no ablation isolates the contribution of the constraints. The central claim is explicitly about empirical performance, especially Mean Rank, so the ambiguity directly controls whether the proposed distance-based scorer is responsible for the reported numbers. This is more load-bearing than the secondary theoretical issues in Section V, which are real but affect the paper's narrative rather than the empirical headline. For instance, Eq. (26) is asserted to follow from nonzero imaginary parts of the relation quaternion, but a counterexample (e.g., Wr = i, Qh = 1, Qt = j) gives equal distances on both sides, so the antisymmetry proof is invalid. Yet even if the theory were fixed, the empirical claim would remain unverifiable without specifying the evaluation protocol for QuatE-D2. The reader's verdict of REJECT is therefore unchanged: the paper should be resubmitted with a precise definition of the type constraints, an ablation showing their effect, and the evaluation code.","tokens_in":14986,"tokens_out":6436,"duration_ms":56517,"concrete_test":"Re-implement QuatE-D exactly as defined by Eqs. (18)-(22) and reproduce QuatE-D1 on WN18RR with the standard filtered protocol. Then add a plausible type-constraint mechanism: use the training set to compute, for each relation, the set of entities observed as head and as tail with that relation, and restrict candidate ranking for each test triple to those type-compatible entities. If the resulting MR and MRR match QuatE-D2 (1050 and 0.483) rather than QuatE-D1 (4206 and 0.468), the headline gains are fully explained by the filter rather than by the scoring function. The authors should also be required to release the evaluation script that produces the QuatE-D2 row, so the entry point of the constraints can be checked directly.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline claim that QuatE-D excels in Mean Rank rests on the QuatE-D2 rows of Tables IV and V. The only description given is the table footnote 'WITH TYPE CONSTRAINTS'; the main text never states what these constraints are, whether they are a learned part of the model, a training-time regularizer, or a test-time candidate filter. The D1-to-D2 deltas have exactly the signature of an evaluation filter: MR improves dramatically (WN18 366 to 160, WN18RR 4206 to 1050, FB15k 45 to 21, FB15k-237 131 to 71) while MRR and Hits@K change only slightly. If the constraints are applied at ranking time by restricting candidates to entities of the relation's type signature, the reported MR gains do not measure the distance-based scoring function; they measure the filter. Since Eqs. (20)-(22) define only the unconstrained model, the empirical case for the central claim is currently not assessable.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes QuatE-D, a knowledge graph embedding model that represents entities and relations as quaternions and scores triples by the Euclidean distance between the Hamilton-rotated head and the tail: phi(h,r,t) = ||Q_h (x) W_r^norm - Q_t||_2 (Eq. 20). Training uses a claimed margin ranking loss with L2 regularization (Eqs. 21-22). The authors argue that quaternion non-commutativity gives antisymmetry and composition, and that inversion is handled by quaternion conjugation, with RotatE as a special case. On WN18, FB15k, WN18RR, and FB15k-237, they report filtered link prediction results for 'QuatE-D1' (without type constraints) and 'QuatE-D2' (with type constraints), claiming state-of-the-art or competitive performance, especially in Mean Rank.","tokens_in":15214,"tokens_out":10285,"duration_ms":83188,"significance":"The proposed scoring function is simple and interpretable, and the paper follows standard evaluation protocols on widely used benchmarks, reporting filtered results and hyperparameter searches. If the method as defined by Eq. (20) genuinely delivered the reported gains, a distance-based quaternion model would be a useful baseline. However, the central claim depends on the undefined 'type constraints' variant (D2), and the theoretical support is weakened by algebraic errors in the norm computations and by a loss function that does not implement the described margin ranking. These issues prevent the paper from supporting its conclusions as written.","major_comments":[{"comment":"The headline results for the central claim--'particularly excelling in Mean Rank'--are the QuatE-D2 rows of Tables IV and V, which are labeled only '[with type constraints]' in the table footnotes. The main text never defines these type constraints: they do not appear in the model description (Section IV), the loss (Eq. 22), or the evaluation protocol (Section VI-B). The D1-to-D2 deltas (e.g., WN18RR MR 4206 to 1050; FB15k-237 MR 131 to 71 while MRR changes only from 0.425 to 0.443) are consistent with an evaluation-time filter on candidate entities, rather than a learned model property. Because it is unknown whether the constraints use test-time information, the reported MR gains do not measure the distance-based scoring function of Eq. (20). For the defined model QuatE-D1, the MR results are not competitive (e.g., WN18RR 4206 vs. RotatE 3277; FB15k 45 vs. RotatE 40), so the paper's central empirical claim is not interpretable from the evidence provided.","section":"Tables IV and V; Section VI-E"},{"comment":"The loss in Eq. (21) does not implement the margin ranking objective described in the text. With Y=1 for positives, the positive-triple term is max(0, gamma + phi(h,r,t)), which pushes the distance toward zero; with Y=-1, the negative term is max(0, gamma - phi(h',r,t')), which pushes the negative distance above gamma. There is no pairwise term comparing phi_pos and phi_neg, so the claimed property phi_pos + gamma <= phi_neg is not enforced. In addition, the summation indices in Eq. (22) over '(h,r,t) union (h',r,t')' do not specify how positive and negative samples are paired. Since this loss defines the model whose results are reported, the method as stated is not reproducible.","section":"Section IV, Eq. (21)-(22)"},{"comment":"The proofs of antisymmetry and inversion rely on the identity |A (x) B| = |A| |B| applied to quaternion vectors, writing for instance ||Q'_h||^2 = ||Q_h||^2 ||W^norm_r||^2. For a vector of k quaternions, the squared Euclidean norm of the component-wise Hamilton product is sum_i |q_i w_i|^2, which is not equal to (sum_i |q_i|^2)(sum_i |w_i|^2) in general. The identity holds for a single quaternion, not for the vector norm used in Eq. (20). This invalidates the norm expansions in Eqs. (27), (28), (33), (34), and (38), on which the claimed theoretical properties rest.","section":"Section V-B, Eqs. (27)-(28) and (33)-(38)"},{"comment":"The inversion proof does not actually involve an inverse relation. Equation (32) equates distances computed with the same relation embedding W^norm_r on both sides; it asserts ||Q_h (x) W^norm_r - Q_t||_2 = ||Q_t (x) W^norm_r - Q_h||_2, which is just the symmetry of the Euclidean norm for the two points Q_h (x) W^norm_r and Q_t. To prove inversion one would need to relate the embeddings of r1 and r2 (e.g., W_{r2} as the conjugate of W_{r1}) and show that the distance equality follows from that relation; the manuscript never introduces W_{r1} and W_{r2}. Consequently, the inversion property is unsupported.","section":"Section V-B2, Inversion"},{"comment":"The antisymmetry proof claims that Eq. (26) holds whenever the imaginary parts of W are nonzero, but the argument is incomplete: the expansion (29)-(31) only asserts that the two inner products are different because quaternion multiplication is non-commutative. Even if the inner products differ, it does not follow that the corresponding squared distances differ, since the quadratic terms may compensate. Moreover, the intended conclusion--that the scoring function can distinguish r(x,y) from r(y,x)--requires the inequality to hold for the distance in Eq. (20), not merely for generic quaternion non-commutativity. No counterexample or universal argument is supplied, so the property is not established.","section":"Section V-B1, Antisymmetry"}],"minor_comments":[{"comment":"The norm is defined as ||q|| = q q* = q0^2 + q1^2 + q2^2 + q3^2; this is the squared norm, not the norm. The correct expression is ||q|| = sqrt(qq*) = sqrt(q0^2 + q1^2 + q2^2 + q3^2).","section":"Section III, Eq. (13)"},{"comment":"The text says 'provided in Table 6' but the table is numbered Table III.","section":"Section VI-D, Table III"},{"comment":"The figure is titled 'Impact of Embedding Dimension on Triple Classification Performance,' yet no triple classification task or protocol is described in Section VI; the figure appears to be based on link prediction or a different evaluation.","section":"Figure 1"},{"comment":"The sentence 'DistMult exhibits higher negative sampling rates, which may affect its overall ranking performance' is unclear; negative sampling rates are not reported for baselines.","section":"Section VI-E"},{"comment":"The relation indices in Eq. (39) appear inconsistent: the statement r1 = r2 (compose) r3 with the preceding conditions z = r1(y), y = r3(x), x = r2(y) would imply different composition order; please check the notation.","section":"Section V-B3, Eq. (39)"},{"comment":"There are recurring typos in model names (e.g., 'DuatE' for DualE) and inconsistent use of 'QuatE-D1/QuatE-D2' and 'Quat-D 2' in Table VI; these should be harmonized.","section":"Throughout"}],"recommendation":"reject","confidential_remarks":"The paper appears to be a preliminary draft; the undefined 'type constraints' with large gains is a red flag for evaluation leakage, and the theoretical errors are load-bearing. I recommend rejection. If the authors can clarify the constraints, replace the loss with a standard margin ranking objective, and rerun the experiments for the unconstrained model with the corrected evaluation, a revised submission could be reconsidered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: the paper has a clear, small idea—replace QuatE's inner-product score with Euclidean distance on the rotated head—and it correctly shows that in the complex plane this reduces to RotatE. That part is solid. But the headline Mean Rank numbers come from a variant labeled only \"with type constraints,\" which is never defined, and the closest reading is that it's a test-time candidate filter. That makes the main empirical claim uninterpretable.\n\nWhat is genuinely new: the specific combination of quaternion Hamilton product and distance scoring, plus a type-constrained variant. It's a legitimate extension of an existing program, not a conceptual leap. The paper is easy to follow, and the RotatE reduction is shown explicitly. That is the best part.\n\nThe problems, in order of severity:\n\n1. QuatE-D2, the variant with the best MR results (160 on WN18, 21 on FB15k, 1050 on WN18RR, 71 on FB15k-237), is described nowhere. The table footnotes just say \"with type constraints.\" If those constraints are applied at ranking time to restrict candidates to entities of the relation's type, then the MR gains do not measure the distance-based scorer at all. The D1-to-D2 deltas have exactly that signature: MR improves by a factor of 2-4 while MRR and Hits@K barely move. The paper must specify what these constraints are, where they enter, and whether they use test-time information.\n\n2. The loss function in Eq. (21) is not the margin ranking loss described in the text. For a positive triple it just minimizes phi; for a negative triple it pushes phi above gamma. It never compares the positive score to the negative score. The described pairwise margin has phi_pos + gamma <= phi_neg, which would be max(0, gamma + phi_pos - phi_neg). As written, training does not do what the paper claims.\n\n3. The theoretical properties are shaky. The inversion proof uses |A⊗B|=|A||B| and then asserts an inner-product equality that doesn't follow. The antisymmetry proof claims non-commutativity alone ensures the distances differ; that's not true for all quaternions. These are fixable, but they're not established.\n\nNo code or data is released, so the results can't be checked independently.\n\nWho gets value: someone working on quaternion KGE might use the distance-scoring variant as a baseline once the type constraints are clarified. As written, I wouldn't trust the numbers.\n\nRecommendation: desk-reject for now. The idea is not bad, but the central empirical claim is uninterpretable and the loss is mis-specified. Tell the authors to fix those two things, define the type constraints, and release code. Then it's worth a look.","headline":"QuatE-D is a clear, small idea that reduces to RotatE in the complex plane, but its headline MR numbers come from an undefined 'type constraints' variant, and its loss equation is not the margin ranking loss the text describes.","tokens_in":15688,"tokens_out":4770,"would_cite":false,"duration_ms":42058,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"QuatE-D replaces the inner-product scoring of quaternion knowledge-graph embeddings with Euclidean distance, and reports lower Mean Rank on four link-prediction benchmarks.","keywords":["knowledge graph embedding","quaternion embeddings","distance-based scoring","Hamilton product","link prediction","mean rank","WN18RR","FB15k-237"],"falsifier":"An independent reimplementation that defines the type constraints and applies them identically to all baselines should reproduce MR 160 on WN18 and 71 on FB15k-237; if the constraints are only applied during ranking of candidate tails, the gap between QuatE-D1 and QuatE-D2 would expose them as evaluation-time filtering rather than learned geometry.","tokens_in":14807,"feed_emoji":"🔗","tokens_out":6386,"duration_ms":51700,"temperature":0.7,"pith_summary":"QuatE-D is a quaternion knowledge-graph embedding model whose scoring function measures Euclidean distance between the relation-rotated head embedding and the tail embedding, $\\phi(h,r,t)=\\|Q_h \\otimes W_r^\\triangleleft - Q_t\\|_2$, instead of the inner-product score used by QuatE. The paper argues that this distance-based score is more geometrically interpretable and flexible, and that it captures symmetry, antisymmetry, inversion, and composition through quaternion algebra. On the WN18, FB15k, WN18RR, and FB15k-237 benchmarks, the model reports competitive or best results, with its most consistent gains in Mean Rank. The central message is that the choice of scoring geometry, not just the embedding algebra, is a first-order factor in link prediction.","feed_headline":"Distance-based quaternion scoring cuts link-prediction Mean Rank","feed_subtitle":"Euclidean distance instead of inner products lifts quaternion link prediction on four benchmarks.","key_machinery":"The load-bearing object is the normalized relation quaternion $W_r^\\triangleleft$ together with the Hamilton product $Q_h \\otimes W_r^\\triangleleft$ followed by Euclidean distance $\\|Q_h \\otimes W_r^\\triangleleft - Q_t\\|_2$. Normalization forces $\\|W_r^\\triangleleft\\|=1$, which makes Hamilton multiplication norm-preserving; associativity yields composition, non-commutativity gives directionality, and zero imaginary parts give symmetry. The same machinery reduces to RotatE when the quaternions are restricted to the complex plane, making RotatE a special case of QuatE-D.","core_discovery":"The central claim is that replacing the Hermitian inner product of the quaternion model QuatE with the Euclidean distance between a Hamilton-rotated head and the tail gives a more general and more accurate scoring rule. The paper defines $Q'_h = Q_h \\otimes W_r^\\triangleleft$ with $W_r^\\triangleleft = W_r / |W_r|$, scores a triple as $\\|Q'_h - Q_t\\|_2$, and trains with a margin ranking loss plus $\\ell^2$ regularization. Under the restriction to complex quaternions the score reduces exactly to RotatE's $\\|h \\circ r - t\\|$, so QuatE-D is presented as the quaternion generalization of rotation-based distance scoring. The reported tables show QuatE-D2, the variant with type constraints, achieving the lowest Mean Rank on all four datasets (160 on WN18, 21 on FB15k, 1050 on WN18RR, 71 on FB15k-237) and the best or tied-best MRR and Hits on WN18RR and FB15k-237. The type-constraint mechanism itself is not defined in the paper.","pith_inferences":["The unnamed 'type constraints' behind QuatE-D2's headline numbers may be doing some of the work; a re-run with the constraints removed or fully specified would tell whether the distance scorer alone delivers the gains.","The same substitution of inner product with Euclidean distance could be applied inside other quaternion and dual-quaternion models, and the paper's own future-work section suggests this; the transfer is plausible because the score only changes how the rotated head is compared to the tail.","Because the model reduces to RotatE, a controlled dimension-for-dimension comparison against RotatE with identical negative sampling would isolate the benefit of the fourth quaternion component; the current tables do not report such a match."],"forward_implications":["On WN18, QuatE-D2 reports Mean Rank 160, below RotatE's 184 and QuatE's 388, with Hit@3 0.956 and Hit@10 0.960.","On FB15k-237, QuatE-D2 reports the best compared MR (71), MRR (0.443), Hit@1 (0.344), Hit@3 (0.487), and Hit@10 (0.638).","Because the score reduces to RotatE's when quaternions are confined to the complex plane, QuatE-D is a strict generalization of rotation-based distance scoring rather than a separate family.","The theoretical section argues that non-zero imaginary parts in $W_r$ give antisymmetric scores, conjugate relations give inverse scores, and associativity gives composition, so the distance geometry is claimed to support the four classic relational patterns."],"supporting_citations":[{"why":"Provides the TransE baseline, the WN18 and FB15k datasets, and the filtered evaluation protocol used throughout.","marker":"[7]"},{"why":"Supplies the RotatE baseline whose rotation-distance score QuatE-D extends and to which QuatE-D reduces on the complex plane.","marker":"[8]"},{"why":"Introduces the quaternion inner-product model QuatE that QuatE-D directly replaces with distance scoring.","marker":"[9]"},{"why":"Dual quaternion model DualE is the strongest quaternion-family competitor, with higher parameter count, on MRR and Hits.","marker":"[23]"},{"why":"Introduces the WN18RR benchmark and supplies the ConvE baseline and several reproduced comparison numbers.","marker":"[31]"},{"why":"Introduces the FB15k-237 benchmark, needed to test without inverse-relation shortcuts.","marker":"[41]"},{"why":"HolE is a baseline and the stated source of several WN18 and FB15k results in the comparison tables.","marker":"[22]"}],"fun_headline_variants":["QuatE-D: Euclidean distance beats inner product for quaternion KGE","Distance scoring in quaternion space reduces Mean Rank","Type-constrained QuatE-D achieves lowest Mean Rank on four benchmarks","QuatE-D generalizes RotatE with distance-based quaternion scoring"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The best reported results depend on a 'type constraints' variant that is never defined, so the distance-based scoring function may not be what produces the headline Mean Rank gains.","fun_headline_variants_meta":{"raw":{"variants":["QuatE-D: Euclidean distance beats inner product for quaternion KGE","Distance scoring in quaternion space reduces Mean Rank","Type-constrained QuatE-D achieves lowest Mean Rank on four benchmarks","QuatE-D generalizes RotatE with distance-based quaternion scoring"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001202,"raw_usage":{"total_tokens":4930,"prompt_tokens":899,"completion_tokens":4031,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":515,"completion_tokens_details":{"reasoning_tokens":3955}},"tokens_in":515,"tokens_out":4031,"duration_ms":25104,"temperature":1.0,"reasoning_tokens":3955,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:06:12.551289+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"An independent reimplementation that defines the type constraints and applies them identically to all baselines should reproduce MR 160 on WN18 and 71 on FB15k-237; if the constraints are only applied during ranking of candidate tails, the gap between QuatE-D1 and QuatE-D2 would expose them as evaluation-time filtering rather than learned geometry.","supporting_citations":[{"cited_title":"Translating embeddings for modeling multi-relational data,","cited_arxiv_id":null,"evidence_quote":"Provides the TransE baseline, the WN18 and FB15k datasets, and the filtered evaluation protocol used throughout."},{"cited_title":"Dual quaternion knowledge graph embeddings,","cited_arxiv_id":null,"evidence_quote":"Dual quaternion model DualE is the strongest quaternion-family competitor, with higher parameter count, on MRR and Hits."},{"cited_title":"Convolutional 2d knowledge graph embeddings,","cited_arxiv_id":null,"evidence_quote":"Introduces the WN18RR benchmark and supplies the ConvE baseline and several reproduced comparison numbers."},{"cited_title":"Observed versus latent features for knowl- edge base and text inference,","cited_arxiv_id":null,"evidence_quote":"Introduces the FB15k-237 benchmark, needed to test without inverse-relation shortcuts."},{"cited_title":"Holographic embeddings of knowledge graphs,","cited_arxiv_id":null,"evidence_quote":"HolE is a baseline and the stated source of several WN18 and FB15k results in the comparison tables."}],"review_version":1}