Powers of Magnetic Graph Matrix: Fourier Spectrum, Walk Compression, and Applications

David F. Gleich; Pan Li; Yinan Huang

REVIEW 2 major objections 3 minor 26 references

Powers of Magnetic Graph Matrix: Fourier Spectrum, Walk Compression, and Applications

T0 review · 2 major / 3 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper proves that the $m$-th power of a directed graph's magnetic matrix is exactly the discrete Fourier transform of its $m$-step bidirectional walk counts, so local directed structure can be recovered from a few matrix potentials.

desk verdict The Fourier/walk-profile identity is a genuinely useful bridge, but the claimed exact reconstruction with floor(m/2)+1 potentials is off by one for odd m—the Nyquist potential is dropped. read the letter →

arxiv 2506.07343 v1 pith:2ZNN2M4T submitted 2025-06-09 physics.soc-ph cs.SI

Yinan Huang , David F. Gleich , Pan Li This is my paper

classification physics.soc-phcs.SI MSC 05C5005C2005C38

keywords magneticgraphmatrixdirectednetworkswalkprofilesbidirectionalwalksFouriertransformcompressionnetworkmotifslinkprediction

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper aims to give powers of the magnetic graph matrix a concrete combinatorial meaning. The magnetic matrix $A_q = A e^{i2\pi q} + A^\top e^{-i2\pi q}$ places a complex phase on each directed edge, and the paper proves that its $m$-th power $[A_q^m]_{u,v}$ is exactly the Fourier transform, at frequency $2q$, of the walk profile $\Phi_{u,v}(m,\cdot)$: the counts of $m$-step walks from $u$ to $v$ with a fixed number of forward and backward edges. Because $q$ acts as a Fourier frequency, $\lfloor m/2 \rfloor + 1$ carefully chosen potentials reconstruct every walk profile exactly, and on most real networks far fewer potentials reach about 5% reconstruction error. This matters because powers of a graph matrix drive random-walk, diffusion, and centrality computations; the paper shows the walk-profile view detects directed motifs that earlier magnetic-Laplacian spectral analysis missed and improves directed link prediction.

What carries the argument

The load-bearing object is the walk profile $\Phi_{u,v}(m,k)$, the count of length-$m$ bidirectional walks from $u$ to $v$ with $k$ forward and $m-k$ backward edges, which obeys the recursion $\Phi(m,k) = \Phi(m-1,k-1)A + \Phi(m-1,k)A^\top$. The carrying mechanism is the binomial expansion $A_q^m = \sum_{k=0}^m e^{i2\pi q(2k-m)}\Phi(m,k)$, which rewrites each magnetic matrix power as a discrete Fourier transform of the profile vector at frequency $2q$. The expansion works backward too: evaluating at the $q_j = j/(2(m+1))$ grid and exploiting the symmetry $A^m_q = (-1)^m (A^m_{1/2-q})^*$ reduces the needed count from $m+1$ potentials to $\lfloor m/2 \rfloor + 1$ and reconstructs the profiles by inverse Fourier transform. For the compression claim the operative quantity is the spectral radius $\rho(\hat{A}_q)$ of the random-walk magnetic matrix, whose decay in $q$ the paper measures and correlates with reconstruction error.

What would settle it

A concrete check: on the paper's ogbg-code2 program graphs, which are directed trees with $\rho(\hat{A}_q) \equiv 1$, reconstruction from ten potentials stays far above the 5% threshold; finding any directed graph where $\rho(\hat{A}_q)$ is flat across $q \in [0,1/4]$ yet walk profiles reconstruct below 5% error with about $m/5$ potentials would break the claimed spectral-radius-to-compressibility mechanism, since the paper's measured correlation says that combination should not occur.

Watch

Extended reading notes

Core claim

The central discovery is the identity $[A_q^m]_{u,v} = e^{-i2\pi qm}\sum_{k=0}^m e^{i4\pi qk}\Phi_{u,v}(m,k)$, where $\Phi_{u,v}(m,k)$ counts $m$-step bidirectional walks from $u$ to $v$ with exactly $k$ forward and $m-k$ backward edges. Expanding $(A e^{i2\pi q} + A^\top e^{-i2\pi q})^m$ by the binomial theorem groups matrix products by how many factors are $A$ rather than $A^\top$, and each group is one walk-profile entry, so only the group's phase depends on the potential $q$. The identity makes the collected powers at potentials $q_j = j/(2(m+1))$ an invertible Fourier system: the inverse transform in Eq. (3) recovers all walk profiles exactly from $\lfloor m/2 \rfloor + 1$ potentials, and the same relation holds for the random-walk-normalized matrix $D^{-1}A_q$, where profiles become bidirectional random-walk landing probabilities. On 26 real directed networks the paper finds the profile spectrum concentrated at low frequencies, so the smallest few potentials reconstruct $\hat{\Phi}(m,\cdot)$ to roughly 5% error, and the reconstruction error tracks the decay of the spectral radius $\rho(\hat{A}_q)$.

Load-bearing premise

The load-bearing premise is empirical: real directed networks concentrate their walk-profile energy at low frequencies, with a spectral radius $\rho(\hat{A}_q)$ that shrinks as the potential $q$ grows; the paper demonstrates this on 26 datasets but also shows the premise fails for entire graph classes, since directed trees have $\rho(\hat{A}_q) \equiv 1$ and do not compress.

Editorial extensions

If this is right

Every $m$-step walk profile of a directed graph is exactly recoverable from $\lfloor m/2 \rfloor + 1$ chosen powers $A_q^m$, and powers at any other potential can then be synthesized from Eq. (1).
On the 26 real networks studied, about $m/5$ low potentials give roughly 5% reconstruction error, so a handful of complex matrix powers can stand in for full bidirectional-walk statistics.
Compressibility is governed by how fast $\rho(\hat{A}_q)$ decays with $q$: networks whose spectral radius stays near 1, notably directed trees, resist compression and need the full set of potentials.
Walk profiles and magnetic matrix powers detect fully oriented cycles and cycles with exactly one backward edge, such as feed-forward loops, and provably cannot tell apart cycles whose two or more backward edges are arranged differently.
As link-prediction features, magnetic walk powers match or beat directed, undirected, and random-walk-proximity baselines, even with a single potential.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Beyond the paper: any method built on powers of $A_q$ sees only the number of edge reversals in a walk, never their ordering, so products or correlations of powers at two potentials are a natural next probe for orientation-order information that a single power provably loses.
Beyond the paper: walk profiles count all walks including backtracking; replacing the recursion with a non-backtracking transfer operator would yield an analogous Fourier identity for non-backtracking bidirectional walks, a direction the paper flags for future work.
Beyond the paper: the reported $r = 0.95$ link between $\rho(\hat{A}_q)$ decay and reconstruction error suggests a cheap pre-test for a new network: measure the spectral radius at two potentials and predict whether walk-profile features will compress before computing any high powers.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Let me cut to the chase: the core identity is right and worth knowing, but the main reconstruction theorem is wrong for odd m. The stress-test note checks out. When m is odd, N=m+1 is even, so a real DFT has an independent Nyquist bin at q=1/4. Equation (3) and supplement Eq. (24) omit it. The supplement's Eq. (23) initially includes that term and then asserts it vanishes; the assertion is false. Concretely, A_{1/4}^m = i^m(A-A^T)^m, and i^m A_{1/4}^m is generally a nonzero real skew-symmetric matrix. The counterexample in the stress-test (3-edge graph, m=3, pair (1,3)) gives Phi(3,1)=3 and Phi(3,0)=1 from direct enumeration, while Eq. (3) with only q_0 and q_1 yields 2.5 and 1.5. So the exact number of potentials for odd m is floor(m/2)+2, not floor(m/2)+1. Theorem 1 as stated is false. The experiments use m=50 and m=20, both even, so the numerics survive; but the theorem needs correction, and the paper should be checked for downstream statements that inherit the error.

What is good: the walk-profile/Fourier connection is a clean and self-contained bridge between local combinatorial structure and magnetic spectral methods. Eq. (1) is derived carefully. The compression study is honest—trees are explicitly flagged as hard cases, and the correlation with spectral radius (r=0.95, CI reported) is a solid empirical observation. The motif-detection negative result, though idealized by the no-repeat assumption, is a thoughtful limitation statement. The applications show real promise for local directed-graph feature engineering.

Where it is soft: aside from the theorem bug, the 'm/5 potentials give 5% error' claim is dataset-dependent; the paper mostly says this, which is credit to it. The motif detection capacity argument assumes no repeated nodes, and real walks do repeat; the authors acknowledge this, but it does limit the strength of the claim.

Overall: a solid submission with a real, fixable flaw. I would send it to referees rather than desk-reject, but not accept as is. The reader's ACCEPT missed the Nyquist issue; the stress-test is right. The authors should fix the theorem and verify all derived counts.

Referee Report

2 major / 3 minor

Summary. The paper introduces walk profiles, defined as counts of length-m bidirectional walks indexed by the number of forward edges, and shows that the m-th power of the magnetic adjacency matrix A_q = A e^{i2πq} + A^T e^{-i2πq} is, up to a phase factor, the discrete Fourier transform of these walk profiles evaluated at frequency 2q. Building on this identity, the paper claims that walk profiles can be exactly reconstructed from the magnetic matrix powers at ⌊m/2⌋+1 potentials q_j = j/(2(m+1)), and it provides an explicit inverse formula (Eq. (3), Supplement Eq. (24)). The authors further demonstrate empirically that on many real and synthetic directed networks, a much smaller number of low-frequency potentials yields accurate approximate reconstruction, and they relate this compressibility to the decay of the spectral radius of the random-walk magnetic matrix. Two applications are presented: detecting minimally frustrated directed cycles (including feed-forward loops) using diagonal entries of magnetic matrix powers, and link prediction using magnetic-walk features.

Significance. The central identity connecting magnetic matrix powers to walk profiles is clean, correct, and likely useful: it provides a combinatorial interpretation of powers of a Hermitian asymmetric matrix, gives a probabilistic interpretation for the random-walk normalization, and suggests new spectral tools for directed networks. The paper ships reproducible code, the core derivation is straightforward and machine-checkable, and the empirical claims are supported by 26 networks with uncertainty quantification (e.g., r = 0.95 with 95% CI [0.90, 0.98] in Fig. 4). The honest treatment of limitations—for instance, the impossibility result for motif detection with two or more backward edges (Supplement Theorem 2) and the explicitly dataset-dependent nature of compressibility—strengthens the contribution. However, the exact-reconstruction theorem contains a parity flaw for odd m, which is a load-bearing theoretical claim even though the reported experiments use even m.

major comments (2)

[Eq. (3) and Supplement S.II, Eqs. (23)–(24), Theorem 1] The reconstruction formula is incorrect for odd m because it omits the Nyquist component at q = 1/4. For odd m, N = m+1 is even, so the real sequence Φ(m,·) has an independent real Nyquist Fourier coefficient at q = 1/4. The term A_{1/4}^m = i^m (A−A^T)^m is real for odd m (since (A−A^T)^m is real and i^m is ±i) and is generally nonzero off-diagonal. Supplement Eq. (23) initially includes this term via δ_{m,odd}(−1)^k i^m A_{q_{⌈m/2⌉}}^m, but then asserts it must be zero for odd m; that assertion is false. Concrete counterexample: on the graph with directed edges 1→2, 2→1, 2→3, 3→1, take m=3 and pair (1,3). Direct enumeration gives Φ_{1,3}(3,0)=1 and Φ_{1,3}(3,1)=4, while Eq. (3) with only q_0=0 and q_1=1/8 gives 1.5 and 3.5 respectively; including q_2=1/4 corrects both to 1 and 4. Therefore, the exact number of potentials required for odd m is ⌊m/2⌋+2, not ⌊m/2⌋+1. The statements in the section "Spectral sparsity and compressibility of walk profiles" that "evaluating magnetic matrix powers at ⌊m/2⌋+1 distinct frequencies q is required" and that "choosing any arbitrary ⌊m/2⌋+1 distinct potentials in the range [0, 1/4]" can solve the system are also wrong for odd m. The reported experiments use m=50 and m=20 (even), so the numerical results are not invalidated, but the theorem as stated is incorrect.
[Supplement S.II, Theorem 1] The statement of Theorem 1 is internally inconsistent: the text immediately before it concludes "we only need ⌊m/2⌋+1 many potentials," but the theorem says reconstruction is from (A_{q_0}^m, ..., A_{q_{⌊m/2⌋+1}}^m), which is ⌊m/2⌋+2 potentials. When the parity issue is fixed, the minimal count should be ⌊m/2⌋+1 for even m and ⌊m/2⌋+2 for odd m, and the theorem statement and the surrounding discussion must be aligned accordingly.

minor comments (3)

[Supplement S.II, Eqs. (20)–(22)] The derivation of the symmetry relation contains algebraic typos: the intermediate equalities in Eq. (20) (e.g., "A_q^m = (-1)^{2m} A_q^m = (-1)^m e^{-iπm} A_q^m") are not identities as written. The final symmetry A_q^m = (-1)^m (A_{1/2−q}^m)^* is correct and should be derived directly from Eq. (17) without the erroneous intermediate steps.
[Captions of Figures 3 and 9] The captions state that the reconstruction error "goes to zero when Q reaches ⌊m/2⌋+1." This is correct for the even values m=50 and m=20 used in the figures, but it is not correct for general m, so the captions should either state the parity dependence or restrict the claim to even m.
[Main text, after Eq. (2)] The indexing is ambiguous: the text writes q = (q_0, ..., q_{⌊m/2⌋+1}) with q_j = j/(2(m+1)), but the summation in Eq. (3) runs only to ⌊m/2⌋. Please align the notation so that the intended number of potentials is unambiguous, and correct it to match the parity-dependent minimal count.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the core Fourier identity is derived by direct binomial expansion, the reconstruction uses an un-fitted inverse DFT, and the empirical compression claims are explicitly dataset-dependent.

full rationale

The paper's central identity, Eq. (1), is obtained by expanding (A e^{i2πq} + A^T e^{-i2πq})^m binomially and grouping terms by the number of forward edges k; the walk profile Φ(m,k) is defined as exactly that grouped sum, so the Fourier relationship is an algebraic identity rather than an assumed conclusion. Equation (3) and supplement Eq. (24) are standard inverse discrete Fourier reconstructions using the realness of Φ, with no fitted constants or parameters calibrated to the data; the number of potentials follows from DFT orthogonality, not from a fit. The compression and application sections fit logistic-regression weights, but these are downstream classifiers and no fitted weight enters the derivation of Eq. (1) or Eq. (3), so they do not make the predictions circular. The paper's own caveats—e.g., that program graphs and trees have ρ(A_q) ≡ 1 and poor compressibility, and that reconstruction quality varies across datasets—show that the compressibility claim is an empirical observation, not an input baked into the framework. The only self-citations are contextual references to prior positional-encoding work and are not load-bearing; no uniqueness theorem or ansatz is imported from the authors' earlier papers to force the walk-profile construction. Even the reviewer-identified subtlety about the Nyquist term for odd m is a mathematical correctness issue in the reconstruction formula, not a case of the target result being assumed as an input. The derivation chain is self-contained, and no step reduces by construction to its own inputs.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

No numeric constants are fitted for the central Fourier identity. The classifier weights in the application sections are standard machine-learning fits and do not affect the mathematical claims. Hyperparameters like walk length m=50 and the number of sampled nodes N0=2000 are choices, not fitted parameters. No new physical entities are introduced; walk profiles are mathematical definitions.

assumptions (4)

domain assumption A is a real-valued adjacency matrix (or has real edge weights), so walk profiles are real and the real/imaginary symmetry of the Fourier system reduces the required potentials.
Invoked in Section S.II to derive Eq. (24) and the floor(m/2)+1 reconstruction count; for complex edge weights the count changes.
standard math Binomial expansion and grouping of matrix products in (A e^{i theta} + A^T e^{-i theta})^m is valid; matrix multiplication is associative.
This is the backbone of Eq. (1); no non-standard algebraic structure is assumed.
standard math The DFT Fourier matrix with q_j = j/[2(m+1)] is invertible on the space of real length-(m+1) signals when floor(m/2)+1 potentials are used.
Used for exact reconstruction via Eq. (3); follows from standard Fourier inversion for real signals.
domain assumption The random-walk magnetic matrix hat(A)_q = D^{-1}A_q is used for probabilistic interpretations, requiring D invertible (no isolated nodes) and appropriate row normalization.
Used in the spectral sparsity section and Methods; for graphs with isolated nodes the normalization would need modification.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Powers of Magnetic Graph Matrix: Fourier Spectrum, Walk Compression, and Applications." pith.science (2026). https://pith.science/paper/2ZNN2M4T

@misc{pith2026250607343,
  author       = {Pith},
  title        = {Pith review of: Powers of Magnetic Graph Matrix: Fourier Spectrum, Walk Compression, and Applications},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2ZNN2M4T}},
  note         = {Machine review of arXiv:2506.07343}
}

read the original abstract

Magnetic graphs, originally developed to model quantum systems under magnetic fields, have recently emerged as a powerful framework for analyzing complex directed networks. Existing research has primarily used the spectral properties of the magnetic graph matrix to study global and stationary network features. However, their capacity to model local, non-equilibrium behaviors, often described by matrix powers, remains largely unexplored. We present a novel combinatorial interpretation of the magnetic graph matrix powers through directed walk profiles -- counts of graph walks indexed by the number of edge reversals. Crucially, we establish that walk profiles correspond to a Fourier transform of magnetic matrix powers. The connection allows exact reconstruction of walk profiles from magnetic matrix powers at multiple discrete potentials, and more importantly, an even smaller number of potentials often suffices for accurate approximate reconstruction in real networks. This shows the empirical compressibility of the information captured by the magnetic matrix. This fresh perspective suggests new applications; for example, we illustrate how powers of the magnetic matrix can identify frustrated directed cycles (e.g., feedforward loops) and can be effectively employed for link prediction by encoding local structural details in directed graphs.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

26 extracted references · 23 canonical work pages

[1]

The structure of the nervous system of the nematode caenorhabditis elegans.Philos Trans R Soc Lond B Biol Sci, 314(1165):1–340, 1986

John G White, Eileen Southgate, J Nichol Thomson, Sydney Brenner, et al. The structure of the nervous system of the nematode caenorhabditis elegans.Philos Trans R Soc Lond B Biol Sci, 314(1165):1–340, 1986

work page 1986
[2]

Large-scale mapping of human protein–protein interactions by mass spectrometry.Molecular systems biology, 3(1):89, 2007

Rob M Ewing, Peter Chu, Fred Elisma, Hongyan Li, Paul Taylor, Shane Climie, Linda McBroom-Cerajewski, Mark D Robinson, Liam O’Connor, Michael Li, et al. Large-scale mapping of human protein–protein interactions by mass spectrometry.Molecular systems biology, 3(1):89, 2007

work page 2007
[3]

Drugbank 5.0: a major update to the drugbank database for 2018.Nucleic acids research, 46(D1):D1074–D1082, 2018

David S Wishart, Yannick D Feunang, An C Guo, Elvis J Lo, Ana Marcu, Jason R Grant, Tanvir Sajed, Daniel Johnson, Carin Li, Zinat Sayeeda, et al. Drugbank 5.0: a major update to the drugbank database for 2018.Nucleic acids research, 46(D1):D1074–D1082, 2018

2018
[4]

Modeling polypharmacy side effects with graph convolutional networks.Bioinformatics, 34(13):i457–i466, 2018

Marinka Zitnik, Monica Agrawal, and Jure Leskovec. Modeling polypharmacy side effects with graph convolutional networks.Bioinformatics, 34(13):i457–i466, 2018

work page 2018
[5]

Gene ontology: tool for the unification of biology.Nature genetics, 25(1):25–29, 2000

Michael Ashburner, Catherine A Ball, Judith A Blake, David Botstein, Heather Butler, J Michael Cherry, Allan P Davis, Kara Dolinski, Selina S Dwight, Janan T Eppig, et al. Gene ontology: tool for the unification of biology.Nature genetics, 25(1):25–29, 2000

2000
[6]

Large-scale analysis of disease path- ways in the human interactome

Monica Agrawal, Marinka Zitnik, and Jure Leskovec. Large-scale analysis of disease path- ways in the human interactome. InPacific Symposium on Biocomputing. Pacific Symposium on Biocomputing, volume 23, page 111, 2018

work page 2018
[7]

Automat- ing the construction of internet portals with machine learning.Information Retrieval, 3:127–163, 2000

Andrew Kachites McCallum, Kamal Nigam, Jason Rennie, and Kristie Seymore. Automat- ing the construction of internet portals with machine learning.Information Retrieval, 3:127–163, 2000

work page 2000
[8]

Citeseer: An automatic citation indexing system

C Lee Giles, Kurt D Bollacker, and Steve Lawrence. Citeseer: An automatic citation indexing system. InProceedings of the third ACM conference on Digital libraries, pages 89–98, 1998

work page 1998

Show all 26 references

[9]

Open graph benchmark: Datasets for machine learning on graphs.Advances in neural information processing systems, 33:22118–22133, 2020

Weihua Hu, Matthias Fey, Marinka Zitnik, Yuxiao Dong, Hongyu Ren, Bowen Liu, Michele Catasta, and Jure Leskovec. Open graph benchmark: Datasets for machine learning on graphs.Advances in neural information processing systems, 33:22118–22133, 2020

2020
[10]

Graphs over time: densification laws, shrinking diameters and possible explanations

Jure Leskovec, Jon Kleinberg, and Christos Faloutsos. Graphs over time: densification laws, shrinking diameters and possible explanations. InProceedings of the eleventh ACM SIGKDD international conference on Knowledge discovery in data mining, pages 177–187, 2005

2005
[11]

Bowling alone and trust decline in social network sites

Paolo Massa, Martino Salvetti, and Danilo Tomasoni. Bowling alone and trust decline in social network sites. In2009 Eighth IEEE international conference on dependable, autonomic and secure computing, pages 658–663. IEEE, 2009

2009
[12]

Coleman.Introduction to Mathematical Sociology

James S. Coleman.Introduction to Mathematical Sociology. Free Press, New York, 1964

1964
[13]

Trust management for the semantic web

Matthew Richardson, Rakesh Agrawal, and Pedro Domingos. Trust management for the semantic web. InInternational semantic Web conference, pages 351–368. Springer, 2003

2003
[14]

Signed networks in social media

Jure Leskovec, Daniel Huttenlocher, and Jon Kleinberg. Signed networks in social media. InProceedings of the SIGCHI conference on human factors in computing systems, pages 1361–1370, 2010

2010
[15]

The political blogosphere and the 2004 us election: divided they blog

Lada A Adamic and Natalie Glance. The political blogosphere and the 2004 us election: divided they blog. InProceedings of the 3rd international workshop on Link discovery, pages 36–43, 2005

2004
[16]

Local higher-order graph clustering

Hao Yin, Austin R Benson, Jure Leskovec, and David F Gleich. Local higher-order graph clustering. InProceedings of the 23rd ACM SIGKDD international conference on knowledge discovery and data mining, pages 555–564, 2017. 23

2017
[17]

Air traffic control system command center

Federal Aviation Administration. Air traffic control system command center. http://www. fly.faa.gov/
[18]

Why anchorage is not (that) important: Binary ties and sample selection

Tore Opsahl. Why anchorage is not (that) important: Binary ties and sample selection. online] http://toreopsahl, 2011

2011
[19]

Tpugraphs: A performance prediction dataset on large tensor computational graphs.Advances in Neural Information Processing Systems, 36:70355–70375, 2023

Mangpo Phothilimthana, Sami Abu-El-Haija, Kaidi Cao, Bahare Fatemi, Michael Burrows, Charith Mendis, and Bryan Perozzi. Tpugraphs: A performance prediction dataset on large tensor computational graphs.Advances in Neural Information Processing Systems, 36:70355–70375, 2023

2023
[20]

On random graphs I.Publicationes Mathematicae, 6:290–297, 1959

Paul Erd˝ os and Albert R´ enyi. On random graphs I.Publicationes Mathematicae, 6:290–297, 1959

1959
[21]

Directed scale-free graphs

B´ ela Bollob´ as, Christian Borgs, Jennifer T Chayes, and Oliver Riordan. Directed scale-free graphs. InSODA, volume 3, pages 132–139. Baltimore, MD, United States, 2003

2003
[22]

Link prediction in directed networks utilizing the role of reciprocal links.IEEE Access, 8:28668–28680, 2020

Jinsong Li, Jianhua Peng, Shuxin Liu, Xinsheng Ji, Xing Li, and Xinxin Hu. Link prediction in directed networks utilizing the role of reciprocal links.IEEE Access, 8:28668–28680, 2020

2020
[23]

Pytorch geometric signed directed: a software package on graph neural networks for signed and directed graphs

Yixuan He, Xitong Zhang, Junjie Huang, Benedek Rozemberczki, Mihai Cucuringu, and Gesine Reinert. Pytorch geometric signed directed: a software package on graph neural networks for signed and directed graphs. InLearning on Graphs Conference, pages 12–1. PMLR, 2024

2024
[24]

BioSNAP Datasets: Stanford biomedical network dataset collection

Marinka Zitnik, Rok Sosiˇ c, Sagar Maheshwari, , and Jure Leskovec. BioSNAP Datasets: Stanford biomedical network dataset collection. http://snap.stanford.edu/biodata, aug 2018

2018
[25]

SNAP Datasets: Stanford large network dataset collection

Jure Leskovec and Andrej Krevl. SNAP Datasets: Stanford large network dataset collection. http://snap.stanford.edu/data, 2014

2014
[26]

Growing well-connected graphs

Arpita Ghosh and Stephen Boyd. Growing well-connected graphs. InProceedings of the 45th IEEE Conference on Decision and Control, pages 6605–6611. IEEE, 2006. 24

2006

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

The structure of the nervous system of the nematode caenorhabditis elegans.Philos Trans R Soc Lond B Biol Sci, 314(1165):1–340, 1986

John G White, Eileen Southgate, J Nichol Thomson, Sydney Brenner, et al. The structure of the nervous system of the nematode caenorhabditis elegans.Philos Trans R Soc Lond B Biol Sci, 314(1165):1–340, 1986

work page 1986

[2] [2]

Large-scale mapping of human protein–protein interactions by mass spectrometry.Molecular systems biology, 3(1):89, 2007

Rob M Ewing, Peter Chu, Fred Elisma, Hongyan Li, Paul Taylor, Shane Climie, Linda McBroom-Cerajewski, Mark D Robinson, Liam O’Connor, Michael Li, et al. Large-scale mapping of human protein–protein interactions by mass spectrometry.Molecular systems biology, 3(1):89, 2007

work page 2007

[3] [3]

Drugbank 5.0: a major update to the drugbank database for 2018.Nucleic acids research, 46(D1):D1074–D1082, 2018

David S Wishart, Yannick D Feunang, An C Guo, Elvis J Lo, Ana Marcu, Jason R Grant, Tanvir Sajed, Daniel Johnson, Carin Li, Zinat Sayeeda, et al. Drugbank 5.0: a major update to the drugbank database for 2018.Nucleic acids research, 46(D1):D1074–D1082, 2018

2018

[4] [4]

Modeling polypharmacy side effects with graph convolutional networks.Bioinformatics, 34(13):i457–i466, 2018

Marinka Zitnik, Monica Agrawal, and Jure Leskovec. Modeling polypharmacy side effects with graph convolutional networks.Bioinformatics, 34(13):i457–i466, 2018

work page 2018

[5] [5]

Gene ontology: tool for the unification of biology.Nature genetics, 25(1):25–29, 2000

Michael Ashburner, Catherine A Ball, Judith A Blake, David Botstein, Heather Butler, J Michael Cherry, Allan P Davis, Kara Dolinski, Selina S Dwight, Janan T Eppig, et al. Gene ontology: tool for the unification of biology.Nature genetics, 25(1):25–29, 2000

2000

[6] [6]

Large-scale analysis of disease path- ways in the human interactome

Monica Agrawal, Marinka Zitnik, and Jure Leskovec. Large-scale analysis of disease path- ways in the human interactome. InPacific Symposium on Biocomputing. Pacific Symposium on Biocomputing, volume 23, page 111, 2018

work page 2018

[7] [7]

Automat- ing the construction of internet portals with machine learning.Information Retrieval, 3:127–163, 2000

Andrew Kachites McCallum, Kamal Nigam, Jason Rennie, and Kristie Seymore. Automat- ing the construction of internet portals with machine learning.Information Retrieval, 3:127–163, 2000

work page 2000

[8] [8]

Citeseer: An automatic citation indexing system

C Lee Giles, Kurt D Bollacker, and Steve Lawrence. Citeseer: An automatic citation indexing system. InProceedings of the third ACM conference on Digital libraries, pages 89–98, 1998

work page 1998