{"id":"217a028e-2cae-4ce3-b317-c85d59472666","arxiv_id":"2504.16748","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"FD-GCL is an augmentation-free, negative-free graph contrastive learner that uses two fractional-diffusion encoders with different orders to generate local and global views, reporting state-of-the-art node classification on several heterophilic benchmarks.","lead":"This paper builds an unsupervised graph learning method with two fractional-diffusion encoders; using different orders makes one encoder focus on local detail and the other on global structure, and a simple cosine loss with a collapse penalty replaces augmentations and negative samples. The method reports top accuracy on several heterophilic benchmarks and stays competitive on homophilic ones.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 2's local/global separation is proven for a regime (large τ, fixed m, α2<1, shared input) not used in training; the central claim is not established for the actual model.","rationale":"The reader's weakest_assumption already identified the regime change between Theorem 2 and the trained model. My stress-test sharpens this into the single most load-bearing concern: the paper's central contribution—that fractional order α controls view diversity (local vs global)—is formally justified only by Theorem 2, whose assumptions (large skip interval τ, fixed iteration count m, α2<1, identical input signal, no learned weight matrices) are all violated by the actual FD-GCL pipeline. The theorem's proof also contains an ill-posed coefficient comparison (Theorem 2(c)) of terms with different powers τ^{−jα1} and τ^{−jα2}, so even the formal statement is not cleanly established. The only direct evidence for the local/global separation is qualitative (Figures 1 and 2), and several Table 8 configurations use α1 close to 1, where the predicted contrast is minimal. This does not prove the method is wrong—the empirical results are broad and the ablation in Table 3 supports that distinct α values help—but it does mean the theoretical rationale for the method, which is a claimed novelty, is unsupported in the evaluated regime. The reader's CONDITIONAL verdict already reflects this; my analysis does not move it, hence UNCHANGED. A direct spectral test on learned views would settle whether the local/global separation actually holds in the trained model, which is the crux of the central claim.","tokens_in":24329,"tokens_out":22034,"duration_ms":189636,"concrete_test":"Train FD-GCL on Cora and Squirrel using the exact hyperparameters in Table 8 (α1, α2, T, h). After training, compute Z1 and Z2, project them onto the eigenvectors of the normalized Laplacian L, and compute the fraction of total energy in the low-frequency components (e.g., top 10% of eigenvalues) versus high-frequency components for each view. Check whether, for every trained seed and checkpoint, Z2 has a strictly higher low-frequency energy fraction than Z1, and whether the difference is monotone in α2−α1 across the α1 values in Table 8. Additionally, repeat with α1=α2 to verify the contrast disappears. If the ordering fails or reverses on any dataset (e.g., Arxiv-year with α1=0.99), the central local/global claim is not supported in the trained regime.","verdict_should_be":"UNCHANGED","load_bearing_attack":"FD-GCL's central claim—that orders α1<α2 produce local vs global views—rests on Theorem 2, but the theorem is proven in a regime that the trained model does not use. Theorem 2 fixes the number of skip iterations m and lets the skip interval τ→∞ (Appendix E), with F = −(I−A)Z, α2<1, and the same input signal x for both encoders. The trained model (Section 4.3, Table 8) instead uses finite T with skip interval equal to the discretization step h≈0.1–10, so m=T/h is large and τ is small, exactly opposite the theorem's asymptotics; α2=1 is routinely used; and the two encoders first multiply X by different learnable matrices W1,W2, so the inputs to the FDEs are not the same signal x. The theorem's conclusion therefore does not formally cover the evaluated model. The proof also compares coefficients b_{α1,i,j} and b_{α2,i,j} of different powers τ^{−jα1} and τ^{−jα2} (Theorem 2(c)), an ill-posed comparison as stated. Direct support for the local/global separation is qualitative (t-SNE in Fig. 1, PCA in Fig. 2), not a quantitative spectral check. Without the theorem, the paper's stated rationale for choosing α1<α2 is unverified, and several datasets (Pubmed, Computer, Arxiv-year) use α1 ≥ 0.75, where the claimed local/global contrast is minimal. The concern is not that the method fails empirically; it is that the load-bearing theoretical justification for the view-diversity mechanism does not apply to the model as trained.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FD-GCL, an augmentation-free graph contrastive learning method in which two encoders are governed by fractional-order differential equations (FDEs) with different fractional orders α1 < α2. The central claim is that the lower-order encoder captures local detail while the higher-order encoder captures global structure, providing the distinct views needed for contrastive learning. The model uses a regularized cosine-mean loss without negative samples, and the authors report state-of-the-art or competitive unsupervised node classification on both homophilic and heterophilic benchmarks, together with ablations, complexity analysis, and qualitative visualizations.","tokens_in":24710,"tokens_out":3323,"duration_ms":35126,"significance":"If the local/global view-diversity mechanism is valid, FD-GCL is a simple and scalable contribution to augmentation-free GCL: it removes both data augmentation and negative sampling, is applicable to heterophilic graphs, and is supported by a broad empirical study with public splits, ten seeds, and consistent ablation results. The paper also provides a theoretical analysis via Mittag-Leffler asymptotics, which is commendable in principle. However, the formal theorem that carries the load for the view-diversity claim is proved under assumptions that the trained model does not satisfy, and the proof itself contains a questionable comparison of asymptotic terms. These issues are central rather than cosmetic, so the theoretical justification requires substantial revision even though the empirical results are promising.","major_comments":[{"comment":"The formal view-separation result is proved for a regime that is not the one used in training. Theorem 2 assumes the left fractional derivative, F(W,Z)=-(I-A)Z, a connected graph, 0<α1<α2<1, a shared input signal x for both encoders, a fixed number m of skip iterations, and the asymptotic τ→∞. The trained model instead uses the discretization step h as the skip interval with finite T (so m=T/h is large and τ=h is small), α2=1 is used on every dataset in Table 8, the two encoders apply different learnable matrices W1 and W2 before diffusion, and the real graphs are not checked for connectivity. Consequently, Theorem 2 does not formally cover the evaluated model, and the claimed correspondence α1↔local, α2↔global is not established for FD-GCL as trained. The t-SNE and PCA evidence in Section 4.2 is qualitative and self-referential; a quantitative spectral check on the trained encoders, or a theorem for finite T, is needed to support the mechanism claim.","section":"Appendix E, Theorem 2 vs. Section 4.3 and Table 8"},{"comment":"The proof of part (c) compares coefficients b_{α1,i,j} and b_{α2,i,j} that multiply different powers of τ in the asymptotic expansions: b_{α1,i,j} multiplies τ^{-jα1} while b_{α2,i,j} multiplies τ^{-jα2}. Since the powers themselves differ, the inequality b_{α1,i,j}>b_{α2,i,j} is not a well-posed comparison without first specifying the same τ-dependent term or proving dominance of the whole sums. In addition, the expression after Eq. (12) writes e_{α1}(λ,τ)^k as if the asymptotic expansion of e_{α1} could be exponentiated term-by-term with a single O(1/τ) remainder; this step is not justified. The issue is load-bearing because (c) is what makes the local/global contrast quantitative.","section":"Appendix E, proof of Theorem 2(c)"},{"comment":"The text states that 'for most datasets studied in the paper, we choose α1≤0.1 and α2>0.5', but Table 8 reports α1=0.75 for Pubmed, 0.94 for Computer, 0.9 for Photo, 0.9 for Squirrel, 0.9 for Chameleon, and 0.99 for Arxiv-year. For these configurations the claimed local/global contrast between α1 and α2=1 is minimal, so the theoretical rationale does not apply to a substantial portion of the experimental evaluation. This inconsistency should be resolved either by revising the claim or by explaining why small α2-α1 still yields sufficiently diverse views.","section":"Appendix E, last paragraph vs. Table 8"}],"minor_comments":[{"comment":"The Barlow Twins loss formula contains a typo: the second sum is written with C_ii, but it should be over off-diagonal entries C_ij (or C_ij^2).","section":"Section 5.3 and Appendix F.4"},{"comment":"The legend entry 'Barlow T wins' appears to be a rendering artifact of 'Barlow Twins'; please correct the label in both figures.","section":"Figures 5 and 6"},{"comment":"There is a duplicated word in the sentence about the weight β: 'and and the regularization weight η'.","section":"Appendix F.3"},{"comment":"The definition of c1 and c2 as 'unit directional vectors of their respective dominant components' is informal; specifying the PCA computation (e.g., first principal component of which matrix, after which normalization) would make the regularized loss reproducible.","section":"Section 4.3"},{"comment":"The theorem is stated for the left fractional derivative, but the implementation section does not state which fractional-derivative definition and which numerical scheme from Kang et al. (2024) are used; this gap should be mentioned explicitly.","section":"Appendix E"}],"recommendation":"major_revision","confidential_remarks":"The empirical study is much stronger than the theory: the gains on heterophilic datasets are large and the ablations are mostly consistent. My concern is that the paper's headline mechanism rests on Theorem 2, and the current mismatch between the theorem's assumptions and the trained model, plus the asymptotic-comparison issue in the proof, would be quickly noticed by a theoretically minded reader. The revision should either extend the theorem to the finite-T setting with the actual encoder pipeline or substantially weaken the theoretical claims and replace them with direct quantitative measurements of view diversity on the trained encoders. I would also encourage the authors to release code, since none is mentioned in the manuscript."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read it. The new thing is real: nobody has used fractional-diffusion order as the contrastive view knob before, and the regularized cosmean loss is a simple idea that seems to stop collapse. The experiments are the paper's strength: broad coverage, public splits, ten seeds, and the heterophilic gains are substantial (Squirrel, Chameleon, Texas). The ablation showing alpha1 < alpha2 beats alpha1 = alpha2 supports the central empirical claim.\n\nThe soft spot is that the theory is about a different model than the one trained. Theorem 2 is proven for the left fractional derivative, F = -(I-A)Z, a connected graph, the limit of large skip interval tau with a fixed number m of skip iterations, alpha2 < 1, and the same input signal x for both encoders. The trained model uses finite T and small h, so m is large and tau is small; alpha2 = 1 is used in every dataset in Table 8; and the two encoders start from different learned projections W1X and W2X. So the formal local/global separation does not cover the evaluated architecture. The proof also compares coefficients of tau^{-j alpha1} and tau^{-j alpha2} as if they were directly comparable, which is not justified as written. None of this says the method is broken; the empirical ablation is consistent with the intuition. But the paper currently overclaims what the theorem establishes.\n\nSmaller issues: alpha1 is tuned per dataset on validation labels, so the method is not fully parameter-free; no code is released. The paper's own limitations section admits manual tuning is a bottleneck.\n\nVerdict: this deserves peer review. The empirical contribution is strong enough that a serious referee should engage. The theory section needs either a theorem that matches the trained regime, a precise statement of the gap, and a fix to the coefficient comparison, or an honest reframing of the local/global story as a heuristic supported by ablations. Code would help.\n\nWho's this for: graph contrastive learning and graph diffusion researchers. I'd bring it to reading group, and would cite it if I worked in GCL.","headline":"The empirical recipe is strong and the heterophilic gains look real, but the paper's theoretical story is proven for a different regime than the one actually trained, so the local/global view claim is currently an intuition supported by ablations rather than a theorem covering the model.","tokens_in":25257,"tokens_out":2363,"would_cite":true,"duration_ms":24762,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proposes that two fractional-diffusion encoders with different derivative orders produce local and global views, and that contrasting these views yields strong unsupervised node classification without augmentations or negative…","keywords":["fractional differential equations","graph neural diffusion","graph contrastive learning","augmentation-free learning","heterophilic graphs","node classification","spectral graph theory","dimension collapse"],"falsifier":"Take a connected graph, fix $F(W, Z) = -(I-A)Z$, and compute the Fourier coefficients of the two encoders' outputs at the finite diffusion time $T$ used in training. If the ratio of the small-order coefficients to the large-order coefficients does not increase with graph frequency, meaning the small-order encoder does not retain relatively more high-frequency energy, the claimed local/global view separation fails. A second check runs the same comparison on a graph with several disconnected components to test whether connectivity is essential.","tokens_in":24142,"feed_emoji":"🕸️","tokens_out":11243,"duration_ms":89527,"temperature":0.7,"pith_summary":"This paper proposes a contrastive learning method for graphs that needs neither data augmentation nor negative samples. It uses two learnable encoders whose feature evolution is governed by fractional differential equations with different derivative orders $\\alpha_1 < \\alpha_2$. The claim is that the small-order encoder retains local, high-frequency detail while the large-order encoder yields global, smooth structure, so the two outputs form naturally distinct views of the same graph. Training pulls the views together with a mean-cosine loss plus a term that penalizes the angle between their dominant directions, preventing dimension collapse. On node-classification benchmarks, including heterophilic graphs where augmentation-based methods struggle, the method reports the best accuracy on most datasets.","feed_headline":"One fractional-order parameter creates two contrastive graph views","feed_subtitle":"Small and large derivative orders yield local and global views, enabling learning without augmentations or negatives.","key_machinery":"The load-bearing object is the fractional derivative $D_t^{\\alpha}$, an integral-based operator that gives the feature evolution memory of its own past. Solving the fractional diffusion equation mode-by-mode in the eigenbasis of the normalized graph Laplacian turns the dynamics into a scalar fractional relaxation per frequency, whose solution is expressed through the Mittag-Leffler function $e_\\alpha(\\lambda, t) = \\sum_{n \\geq 0} (-1)^n \\lambda^n t^{\\alpha n} / \\Gamma(\\alpha n + 1)$. An asymptotic expansion of this function yields the comparison of Fourier coefficients that separates local from global information. The second device is the regularized cosmean loss $L(Z_1, Z_2) = L_0(Z_1, Z_2) + \\eta |\\langle c_1, c_2\\rangle|$, where $c_1, c_2$ are the unit direction vectors of the dominant components of the two views; the penalty keeps the views from collapsing into a single representation.","core_discovery":"The central discovery is a spectral separation principle: with fractional-order diffusion, the derivative order $\\alpha$ controls how quickly graph-frequency components are damped. For $0 < \\alpha_1 < \\alpha_2 \\leq 1$ and large diffusion time, the features from order $\\alpha_2$ are dominated by smooth components corresponding to small eigenvalues of the normalized Laplacian, while the features from order $\\alpha_1$ retain more energy in spiky high-frequency components; the difference grows with $\\alpha_2 - \\alpha_1$. The paper proves this for the left fractional derivative, linear diffusion $F(W, Z) = -(I-A)Z$, a connected graph, and the asymptotic regime $\\tau \\to \\infty$ with a fixed number of skip iterations, by showing that the Fourier coefficients satisfy $c_{\\alpha_l, i}(T) = \\left(\\sum_{j=0}^{n_l} b_{\\alpha_l, i, j} \\tau^{-j \\alpha_l} + O(1/\\tau)\\right) c_i$ with positive coefficients $b$ that decrease in the frequency index $i$ and are larger for the smaller order. On this basis, the paper argues that the two encoders give local and global views, and that a regularized mean-cosine loss can contrast them without negatives.","pith_inferences":["The spectral separation principle is more general than the fractional-diffusion solver used here: any pair of encoders whose frequency responses decay at different rates should produce similarly useful contrastive views, so the mechanism could be transferred to cheap linear filters or polynomial spectral filters.","Because the formal theorem assumes a connected graph, disconnected or near-disconnected graphs may need a spectral shift or regularization; this is a testable boundary of the method, not a claim made in the paper.","The regularized-cosmean idea of penalizing the alignment of dominant component directions could be lifted into other non-contrastive losses to stabilize training, independent of fractional diffusion.","The optimal $\\alpha_1$ varies strongly across datasets (0.001 on Roman-empire, 0.99 on Arxiv-year), which suggests that data-driven or adaptive order selection, which the paper names as future work, may be where the next gains come from."],"forward_implications":["Unsupervised node classification improves on heterophilic graphs; the paper reports relative gains of 9.87% on Squirrel, 5.2% on Chameleon, and 6.22% on Texas over the second-best baseline.","The method removes two design choices at once: no data augmentations and no negative samples, so the contrastive pipeline reduces to choosing two orders $\\alpha_1, \\alpha_2$ and the diffusion time $T$.","Separating the orders matters: ablations with $\\alpha_1 = \\alpha_2$ underperform configurations with $\\alpha_1 < \\alpha_2$, and larger $\\alpha_2 - \\alpha_1$ generally helps, in line with the theorem's prediction.","The regularized cosmean loss stays stable over training epochs, whereas plain Euclidean, VICReg, Barlow Twins, and unregularized cosmean losses degrade; the regularization is credited with preventing dimension collapse.","With the basic GRAND-type diffusion, the FDE encoders run in $O(EC + E \\log E + N)$ per epoch with fast convolution, and the paper reports per-epoch training times on the large Ogbn-Arxiv graph far below the compared augmentation-free baselines."],"supporting_citations":[{"why":"Supplies the GRAND diffusion $F(W, Z) = -(I-A)Z$ and the skip-connection time discretization that the FD-GCL encoders build on.","marker":"Chamberlain et al. (2021b)"},{"why":"Introduces fractional-order graph neural diffusion and the numerical FDE solver used to train the encoders.","marker":"Kang et al. (2024)"},{"why":"Provides the left fractional derivative definition and the Mittag-Leffler solution formula used in the proof of Theorem 2.","marker":"Stinga (2023)"},{"why":"Supplies the asymptotic expansion of the Mittag-Leffler function that yields the comparison of Fourier coefficients in the proof.","marker":"Erdelyi et al. (1955)"},{"why":"Defines graph Fourier analysis, namely Laplacian eigenvectors as graph frequencies, used to state the local/global separation.","marker":"Shuman et al. (2013)"},{"why":"GraphACL provides the strongest augmentation-free baseline results on several heterophilic datasets and is the main comparison FD-GCL reports against.","marker":"Xiao et al. (2023)"},{"why":"PolyGCL provides a spectral-filtering contrastive baseline that FD-GCL outperforms on most heterophilic node-classification tasks.","marker":"Chen et al. (2024)"}],"fun_headline_variants":["Fractional orders create contrastive views without negatives","Spectral separation yields free graph contrastive learning","One diffusion process, two orders, no negatives needed","Simple fractional diffusion powers augmentation-free graph learning","Tuning derivative order gives local and global graph views"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole method rests on the assumption that the local/global distinction proven for a connected graph, a purely linear diffusion, the left fractional derivative, and asymptotically large diffusion time still holds in the finite, learned, possibly disconnected settings where the model is actually trained.","fun_headline_variants_meta":{"raw":{"variants":["Fractional orders create contrastive views without negatives","Spectral separation yields free graph contrastive learning","One diffusion process, two orders, no negatives needed","Simple fractional diffusion powers augmentation-free graph learning","Tuning derivative order gives local and global graph views"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000225,"raw_usage":{"total_tokens":1475,"prompt_tokens":968,"completion_tokens":507,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":584,"completion_tokens_details":{"reasoning_tokens":435}},"tokens_in":584,"tokens_out":507,"duration_ms":5496,"temperature":1.0,"reasoning_tokens":435,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:57:26.620117+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a connected graph, fix $F(W, Z) = -(I-A)Z$, and compute the Fourier coefficients of the two encoders' outputs at the finite diffusion time $T$ used in training. If the ratio of the small-order coefficients to the large-order coefficients does not increase with graph frequency, meaning the small-order encoder does not retain relatively more high-frequency energy, the claimed local/global view separation fails. A second check runs the same comparison on a graph with several disconnected components to test whether connectivity is essential.","supporting_citations":[],"review_version":1}