Pith. sign in

REVIEW 4 major objections 4 minor 47 references

UniSAGE claims that a single attribute graph with orthogonal static/dynamic subspaces can unify hierarchical static data and timestamped records, and that its selective aggregation mechanism can reproduce the output of any hyper-structure g

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 14:47 UTC pith:PN45G7TB

load-bearing objection UniSAGE's empirical story is strong and worth engaging; its universal expressiveness theorem is not actually proven and should be dropped or fixed. the 4 major comments →

arxiv 2607.14102 v1 pith:PN45G7TB submitted 2026-05-06 cs.CL

UniSAGE: Unifying Static and Dynamic Attributes with Hyper-Structure

Taoran Fang , Yan Deng , Chunping Wang , Yang Wang , Lei Chen , Yang Yang This is my paper
classification cs.CL
keywords attribute graphstatic attributesdynamic recordshierarchical datagraph neural networksorthogonal subspacesselective semantic aggregationhyper-structure
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper is trying to establish that hierarchical JSON-like data with both fixed attributes and timestamped event logs can be handled by one automatically built attribute graph, with no schema-specific model engineering. It claims that two parameter spaces kept orthogonal to each other let static composition and dynamic reasoning share one semantic space without interfering, and that a lightweight selective aggregation step, SSAgg, has the expressive power of an explicit task-specific hyper-structure (a subgraph that wires selected attributes together). If this is right, a single model can adapt to evolving schemas and asynchronous records, and the reported results — best scores on all benchmark classification and regression tasks, with several improvements over 10% — would follow from that design.

Core claim

The central claim is that static attributes and dynamic records do not need separate modeling pipelines: both become nodes in one rooted attribute graph, with static edges encoding the schema hierarchy and dynamic edges encoding record order. The root embedding is produced by bottom-up propagation, with static and dynamic transformations constrained to orthogonal parameter subspaces so they do not interfere. The model then re-propagates over the combined adjacency using Selective Semantic Aggregation (SSAgg), an attention mechanism with an extra gate in the denominator. The paper proves that SSAgg can in principle reproduce the representation that any graph neural network would produce on an

What carries the argument

The central object is the global attribute graph: a rooted tree whose nodes are static attributes and individual timestamped records, with static links for hierarchy and dynamic links for time. Two orthogonal parameter subspaces, W_s and W_d, constrain static aggregation and dynamic reasoning so they share a semantic space without cross-talk. The carrying mechanism is SSAgg (Selective Semantic Aggregation), which re-propagates messages over the combined adjacency A' = A_s + A_d using attention with a tunable gate exp(lambda) in the denominator; the gate lets SSAgg suppress selected messages, which the authors argue is equivalent to choosing nodes and edges in an explicit hyper-structure.

Load-bearing premise

The proof that SSAgg can mimic any hyper-structure GNN assumes the attention coefficients it produces can be freely assigned to match the target hyper-structure's aggregation weights, but those coefficients are actually generated by a fixed scoring function over the original adjacency matrix, so the claim may fail for coefficient vectors that scoring function cannot realize; the proof also restricts the target GNN to a single-layer GAT with mean pooling rather than 'any graph

What would settle it

Construct a small original graph and a target hyper-structure subgraph, compute the target GNN's root representation, then solve for the SSAgg attention scores required to reproduce it. If the required scores violate the functional form of the scoring function psi (for example, they cannot be produced by any input pair under the shared gate exp(lambda)), Proposition 2 is false for that case.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If the central claim is correct, practitioners can feed raw JSON into one model and get entity embeddings for classification and regression without hand-designed feature extractors or schema-specific pipelines.
  • Static and dynamic information would live in one shared representation space, so asynchronous records with differing lengths and timestamps no longer require alignment or separate encoders.
  • Because SSAgg provably matches the expressive power of explicit hyper-structure GNNs, expensive node selection and edge construction for task-specific subgraphs can be skipped.
  • The reported benchmark and real-world results imply the design transfers across domains — e-commerce, racing, clinical trials, Q&A communities — and to financial risk scoring with roughly 10% relative gains.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The proof of Proposition 2 establishes existence of parameters in an idealized setting; it does not show that gradient descent will actually find those coefficients. One testable extension is to train SSAgg on a small graph to match a fixed hyper-structure GNN output and check whether learning converges to that target.
  • The paper itself notes in Appendix B.3 that the site-success task could not be trained because the generated graphs were too large and memory-prohibitive, so the claim of handling very large schemas is not fully demonstrated at the extreme scale.
  • The graph construction links a dynamic attribute's parent only to its most recent record; this is a lossy summary. A natural extension would keep a compressed representation of the whole sequence, and the two connectivity choices can be compared directly on tasks where older records are believed to matter.
  • The orthogonality constraint is enforced as a soft penalty, not as an exact projection, so the two subspaces may drift during training; reparameterizing the bases to be exactly orthogonal could be tested as an alternative.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. UniSAGE proposes a graph-based model for hierarchical JSON records containing static attributes and timestamped dynamic logs. It converts each entity into a rooted attribute graph (static hierarchy plus temporal chain), learns root embeddings via two orthogonal parameter subspaces, and then applies SSAgg, an attention-based propagation over A'=As+Ad, to emulate task-specific hyper-structures. The paper claims consistent state-of-the-art results on RelBench entity classification/regression and on a real-world financial dataset, with improvements over 10%, and Propositions 1-2 asserting that SSAgg can suppress arbitrary messages and reproduce the output of any GNN on any hyper-structure graph.

Significance. If the empirical results and the theoretical claims were fully supported, the paper would offer a practical, schema-robust approach to an important data modality, and SSAgg would be a useful lightweight alternative to explicit hyper-structure construction. The paper releases code, evaluates on a broad set of RelBench tasks plus an industrial dataset, and includes ablations and complexity analysis, which are genuine strengths. However, the theoretical guarantee is the main weakness: Proposition 2 is not actually proved as stated, and the formal claim currently overreaches. The empirical comparison also lacks variance reporting. The contribution is potentially significant, but it needs a major revision to make the theory sound and the experimental evidence more rigorous.

major comments (4)
  1. [Appendix A.2, Eq. (40)] The proof of Proposition 2 asserts 'there always exists a set of SSAgg parameters θ such that γ_k = λ_k' without deriving it from Eqs. (21)-(23). In SSAgg, γ_k are outputs of a fixed attention scoring function ψ over A'=As+Ad, not free parameters. The proof never constructs ψ nor shows that the required coefficient vector lies in the attainable set. The claim is therefore assumed rather than proved. Moreover, Eq. (39) writes h'_r as a sum over all k∈V, while Eq. (23) only sums over A'(i,r)=1, so a non-neighbor of the root cannot appear in one SSAgg step.
  2. [Section 4.5, Proposition 2; Appendix A.2, Eqs. (36)-(37)] Proposition 2 states 'any graph neural network', but the proof only treats a single-layer GAT followed by mean pooling, with no nonlinearity, multi-head attention, or general architecture. The target f(V',E') is a graph-level pooled vector, whereas SSAgg produces a root node representation h'_r; equating these is not generally meaningful. The theorem as stated is unsubstantiated. The authors should either restrict the statement to the specific architecture they can prove or supply a genuinely general proof.
  3. [Section 4.3 vs. Section 4.5, Eqs. (20)-(23)] If the hyper-structure G' contains edges that are not in A'=As+Ad, SSAgg has no message path to realize them, since Eqs. (20)-(23) only propagate over existing A' edges. Proposition 1 only shows suppression of messages on existing edges, not creation of new edges. The claim in Section 4.5 that SSAgg 'implicitly realize[s] both node selection and structure construction' is therefore unsupported for arbitrary hyper-structures with new edges.
  4. [Tables 1-2; Appendix B.1] The paper reports average results over five random seeds but gives no standard deviations, error bars, or significance tests. This matters for the repeated claim of 'consistently outperforms', especially where margins are small (e.g., rel-hm user-churn: 70.38 vs. 69.91) and where the headline improvement on UserBehavior is about 4.5 AUC points. Without variance information, it is difficult to judge whether the observed gains are stable or within noise.
minor comments (4)
  1. [Appendix A.1, Eq. (33)] The proof of Proposition 1 says the rearrangement is valid 'as long as δ < 1/n'. In fact, for δ ≥ 1/n the right-hand side 1/δ - n is non-positive, making the inequality automatic for any M. The proposition is still true, but the stated condition is inverted and should be corrected.
  2. [Table 2] '54.1259.06' appears to be missing a space or separator between FreeDyG's result and UniSAGE's result; please fix the typesetting.
  3. [Section 4.2, Eq. (11)] The notation span(W) is not defined. It should be clarified as the column space of W, and it would help to state the dimensions of W_s and W_d explicitly after the orthogonality constraint.
  4. [Appendix B.6] Full-dataset comparison is provided only for rel-f1. The authors should clarify whether the sampled RelBench results in Table 1 are representative of full-dataset behavior, or add full-dataset results for at least one more domain.

Circularity Check

1 steps flagged

Proposition 2's proof assumes SSAgg can realize arbitrary aggregation coefficients (Eq. 40), so the central expressiveness guarantee is asserted rather than derived.

specific steps
  1. other [Appendix A.2, proof of Proposition 2 (Eqs. 36–41), especially Eq. (40); compare SSAgg mechanism Eqs. (21)–(23)]
    "Since bothh′r andhHS G′ are linear combinations of{h∗k}, there always exists a set of SSAgg parametersθsuch that: γk =λk for all k∈V (40)"

    The proof's entire content is Eq. (40): it asserts that SSAgg's aggregation coefficients γ_k can be made equal to the target hyper-structure GNN coefficients λ_k. But in the mechanism, γ_k are not free choices: Eq. (23) computes h′_r as a weighted sum over edges of A'=As+Ad using coefficients β produced by a fixed attention scoring function ψ and the shared gate exp(λ) (Eqs. 21–22). The proof neither constructs ψ nor shows the required coefficient vector is attainable; it simply declares the realization exists. It also ignores that SSAgg has no message path for hyper-edges absent from A', even though Section 4.3 says hyper-structure construction may involve 'generating new edges.' The proof further restricts f to a single-layer GAT with mean pooling (Eqs. 36–37) despite claiming 'any graph

full rationale

The empirical evaluation is self-contained and externally grounded: UniSAGE is compared on public RelBench tasks and a real-world dataset against standard baselines, so the benchmark claims do not reduce to fitted inputs or self-citations. No load-bearing self-citation, uniqueness-import, or renaming pattern appears. The circularity is confined to the theoretical guarantee: Proposition 2's proof reduces the existence of SSAgg parameters to the bare assertion γ_k=λ_k, which is exactly the realizability claim to be proved. The paper itself flags that the proof starts with a simplified single-layer GAT/mean-pooling encoder and never discharges that restriction. Because the proposition is advertised as the theoretical justification for SSAgg ('SSAgg is theoretically guaranteed to match the representational power of explicitly constructed hyper-structures'), the central formal claim is partially circular/unsupported; the empirical results remain independent evidence, keeping the score at 6 rather than higher.

Axiom & Free-Parameter Ledger

2 free parameters · 6 axioms · 1 invented entities

The central method is an empirical architecture, so most assumptions are design choices. Key ledger items: soft orthogonality is substituted for an exact constraint; SSAgg's expressiveness proof assumes free attention weights; the graph construction links a parent only to the latest record; and the private dataset prevents external verification.

free parameters (2)
  • λ (SSAgg suppression gate) = selected from {0,1,2}; best 82.57 at λ=1 on rel-f1 driver-dnf
    Controls selective suppression in Eq. 22; tuned on validation, not derived.
  • γ (orthogonality loss weight) = selected from {0.01,0.1,0.5}; best 82.57 at γ=0.10 on rel-f1 driver-dnf
    Balances the orthogonality regularizer in Eq. 26; chosen by validation.
axioms (6)
  • domain assumption Hierarchical JSON data can be represented as a rooted tree with edges from sub-attributes to parents (Sec. 4.1, Eqs. 5–9)
    If real data contains non-tree references or shared sub-structures, this conversion loses information.
  • ad hoc to paper Connecting a parent only to the most recent record d_T preserves the relevant dynamic state (Eqs. 8–9)
    Heuristic to avoid redundant paths; no analysis of information loss from older records on the static path.
  • domain assumption Soft orthogonality penalty (Eq. 16) with finite γ sufficiently enforces W_s^T W_d = 0 and non-interference (Eq. 11)
    The constraint is not exact; no proof that near-orthogonality preserves the claimed shared semantic space.
  • ad hoc to paper SSAgg can realize arbitrary aggregation coefficient vectors γ_k via attention over A'=As+Ad (Prop. 2 proof, Eqs. 39–41)
    Assumes γ_k are freely matchable, but they are outputs of a fixed attention mechanism; this is the circular step.
  • domain assumption Pretrained Sentence-BERT text embeddings provide sufficient node features for all datasets (Sec. 4.4, Eq. 13)
    No fine-tuning or analysis of text-encoder saturation; performance depends on encoder quality.
  • ad hoc to paper A single-layer GAT with mean pooling represents 'any graph neural network' in the SSAgg expressiveness proof (Prop. 2)
    The proof only covers that specific encoder class, not the general GNN family claimed.
invented entities (1)
  • Task-specific hyper-structure G' no independent evidence
    purpose: A conceptual subgraph that would capture static–dynamic interactions beyond the original graph; SSAgg claims to simulate it without construction
    Introduced as a theoretical device; never instantiated in experiments and no falsifiable handle outside the paper.

pith-pipeline@v1.3.0-alltime-deepseek · 19741 in / 12231 out tokens · 121283 ms · 2026-08-02T14:47:48.675395+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of UniSAGE: Unifying Static and Dynamic Attributes with Hyper-Structure." pith.science (2026). https://pith.science/paper/PN45G7TB

@misc{pith2026260714102,
  author       = {Pith},
  title        = {Pith review of: UniSAGE: Unifying Static and Dynamic Attributes with Hyper-Structure},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PN45G7TB}},
  note         = {Machine review of arXiv:2607.14102}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

With the rapid growth of digital data, real-world applications increasingly involve hierarchical information that combines static attributes with dynamic records. Modeling such heterogeneous data in a unified and generalizable manner remains challenging. Existing approaches often rely on extensive manual design, are tightly coupled to specific data schemas, and typically process static and dynamic attributes in isolation, thereby overlooking their implicit interactions. We propose UniSAGE, a unified framework for modeling data with both static and dynamic attributes. UniSAGE constructs a global attribute graph that represents hierarchical and temporal relationships in a unified structure. To ensure representational consistency, it introduces two orthogonal parameter subspaces that jointly support static aggregation and dynamic reasoning within a shared semantic space. Building on these unified representations, UniSAGE further enables task-specific interaction between static and dynamic attributes via a lightweight hyper-structure mechanism. UniSAGE is fully automated, robust to evolving data schemas, and capable of capturing complex cross-attribute dependencies. Extensive experiments on multiple public benchmarks and a real-world financial behavior dataset demonstrate that UniSAGE consistently outperforms existing methods, achieving performance improvements of over 10% on several tasks.

Figures

Figures reproduced from arXiv: 2607.14102 by Chunping Wang, Lei Chen, Taoran Fang, Yan Deng, Yang Wang, Yang Yang.

Figure 1
Figure 1. Figure 1: Overview of UniSAGE. Step 1: Construct a global attribute graph from JSON, with nodes as attributes/records and edges encoding hierarchical (static) and temporal (dynamic) relations. Step 2: Learn unified representations via orthogonal subspaces for non-interfering static aggregation and dynamic reasoning. Step 3: Capture task-specific cross-type interactions through a lightweight hyper-structure mech￾anis… view at source ↗
Figure 2
Figure 2. Figure 2: The ablation results for UniSAGE [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

47 extracted references · 11 linked inside Pith

  1. [1]

    Biases in electronic health record data due to processes within the healthcare system: retrospec- tive observational study.Bmj, 361,

    [Agnielet al., 2018 ] Denis Agniel, Isaac S Kohane, and Griffin M Weber. Biases in electronic health record data due to processes within the healthcare system: retrospec- tive observational study.Bmj, 361,

    2018

  2. [3]

    Proof.For analytical simplicity, we initially assume the hyper-structure graphG ′ = (V ′,E ′)is encoded using a single-layer GAT model followed by a mean pooling opera- tion over all its nodes. The GAT representation for each node j∈ V′ can be expressed as: hHS j = X k∈N(j) αj,kh∗ k ·W HS (36) whereh ∗ k is the initial representation of nodev k,α j,k is t...

    1991

  3. [8]

    Graph neural networks are dynamic program- mers.ArXiv,

    [Dudzik and Velickovic, 2022] Andrew Dudzik and Petar Velickovic. Graph neural networks are dynamic program- mers.ArXiv,

    2022

  4. [10]

    A compre- hensive survey of dynamic graph neural networks: Mod- els, frameworks, benchmarks, experiments and challenges

    [Fenget al., 2024 ] ZhengZhao Feng, Rui Wang, TianXing Wang, Mingli Song, Sai Wu, and Shuibing He. A compre- hensive survey of dynamic graph neural networks: Mod- els, frameworks, benchmarks, experiments and challenges. arXiv preprint arXiv:2405.00476,

    Pith/arXiv arXiv 2024

  5. [13]

    Inductive representation learning on large graphs.Advances in neural information processing sys- tems, 30,

    [Hamiltonet al., 2017 ] Will Hamilton, Zhitao Ying, and Jure Leskovec. Inductive representation learning on large graphs.Advances in neural information processing sys- tems, 30,

    2017

  6. [14]

    Evaluating build- ing performance in healthcare facilities using entropy and graph heuristic theories.Scientific Reports, 12(1):8973,

    [Hassanainet al., 2022 ] Amr A Hassanain, Mohamed AA Eldosoky, and Ahmed M Soliman. Evaluating build- ing performance in healthcare facilities using entropy and graph heuristic theories.Scientific Reports, 12(1):8973,

    2022

  7. [15]

    Approximation capabilities of multilayer feedforward networks.Neural networks, 4(2):251–257,

    [Hornik, 1991] Kurt Hornik. Approximation capabilities of multilayer feedforward networks.Neural networks, 4(2):251–257,

    1991

  8. [17]

    A review of content-based and context-based recommendation systems.International Journal of Emerging Technologies in Learning (iJET), 16(3):274–306,

    [Javedet al., 2021 ] Umair Javed, Kamran Shaukat, Ibrahim A Hameed, Farhat Iqbal, Talha Mahboob Alam, and Suhuai Luo. A review of content-based and context-based recommendation systems.International Journal of Emerging Technologies in Learning (iJET), 16(3):274–306,

    2021

  9. [21]

    Semi-supervised classification with graph convolutional networks.arXiv preprint arXiv:1609.02907,

    [Kipf, 2016] TN Kipf. Semi-supervised classification with graph convolutional networks.arXiv preprint arXiv:1609.02907,

    Pith/arXiv arXiv 2016

  10. [23]

    Ontological user profiling in recommender systems.ACM Transactions on Information Systems (TOIS), 22(1):54–88,

    [Middletonet al., 2004 ] Stuart E Middleton, Nigel R Shad- bolt, and David C De Roure. Ontological user profiling in recommender systems.ACM Transactions on Information Systems (TOIS), 22(1):54–88,

    2004

  11. [27]

    Sentence-bert: Sentence embeddings using siamese bert-networks

    [Reimers and Gurevych, 2019] Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics, 11

    2019

  12. [29]

    Temporal graph networks for deep learning on dynamic graphs.arXiv preprint arXiv:2006.10637,

    [Rossiet al., 2020 ] Emanuele Rossi, Ben Chamberlain, Fab- rizio Frasca, Davide Eynard, Federico Monti, and Michael Bronstein. Temporal graph networks for deep learning on dynamic graphs.arXiv preprint arXiv:2006.10637,

    Pith/arXiv arXiv 2020

  13. [32]

    Freedyg: Frequency enhanced continuous-time dynamic graph model for link prediction

    [Tianet al., 2024 ] Yuxing Tian, Yiyan Qi, and Fan Guo. Freedyg: Frequency enhanced continuous-time dynamic graph model for link prediction. InThe twelfth interna- tional conference on learning representations,

    2024

  14. [33]

    Graph attention networks.arXiv preprint arXiv:1710.10903,

    [Veliˇckovi´cet al., 2017 ] Petar Veliˇckovi´c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks.arXiv preprint arXiv:1710.10903,

    Pith/arXiv arXiv 2017

  15. [34]

    Review on mining data from multiple data sources.Pattern Recognition Letters, 109:120–128,

    [Wanget al., 2018 ] Ruili Wang, Wanting Ji, Mingzhe Liu, Xun Wang, Jian Weng, Song Deng, Suying Gao, and Chang-an Yuan. Review on mining data from multiple data sources.Pattern Recognition Letters, 109:120–128,

    2018

  16. [35]

    Graph learning based recommender systems: A review.arXiv preprint arXiv:2105.06339,

    [Wanget al., 2021 ] Shoujin Wang, Liang Hu, Yan Wang, Xi- angnan He, Quan Z Sheng, Mehmet A Orgun, Longbing Cao, Francesco Ricci, and Philip S Yu. Graph learning based recommender systems: A review.arXiv preprint arXiv:2105.06339,

    Pith/arXiv arXiv 2021

  17. [36]

    A com- prehensive survey on graph neural networks.IEEE trans- actions on neural networks and learning systems, 32(1):4– 24,

    [Wuet al., 2020 ] Zonghan Wu, Shirui Pan, Fengwen Chen, Guodong Long, Chengqi Zhang, and Philip S Yu. A com- prehensive survey on graph neural networks.IEEE trans- actions on neural networks and learning systems, 32(1):4– 24,

    2020

  18. [37]

    A survey of data representation for multi- modality event detection and evolution.Applied Sciences, 12(4):2204,

    [Xiaoet al., 2022 ] Kejing Xiao, Zhaopeng Qian, and Biao Qin. A survey of data representation for multi- modality event detection and evolution.Applied Sciences, 12(4):2204,

    2022

  19. [38]

    Inductive representation learning on temporal graphs.arXiv preprint arXiv:2002.07962,

    [Xuet al., 2020a ] Da Xu, Chuanwei Ruan, Evren Kor- peoglu, Sushant Kumar, and Kannan Achan. Inductive representation learning on temporal graphs.arXiv preprint arXiv:2002.07962,

    Pith/arXiv arXiv 2002

  20. [39]

    How neural networks extrapolate: From feedforward to graph neural networks.ICLR,

    [Xuet al., 2021 ] Keyulu Xu, Mozhi Zhang, Jingling Li, Si- mon S Du, Ken-ichi Kawarabayashi, and Stefanie Jegelka. How neural networks extrapolate: From feedforward to graph neural networks.ICLR,

    2021

  21. [40]

    Graph neural networks are inherently good generalizers: Insights by bridging gnns and mlps

    [Yanget al., 2022 ] Chenxiao Yang, Qitian Wu, Jiahua Wang, and Junchi Yan. Graph neural networks are inherently good generalizers: Insights by bridging gnns and mlps. ArXiv, abs/2212.09034,

    Pith/arXiv arXiv 2022

  22. [41]

    How graph neural networks learn: Lessons from training dynamics in function space

    [Yanget al., 2023 ] Chenxiao Yang, Qitian Wu, David Wipf, Ruoyu Sun, and Junchi Yan. How graph neural networks learn: Lessons from training dynamics in function space. ArXiv,

    2023

  23. [42]

    Dynamic graph representation learning with neural networks: A survey.Ieee Access, 12:43460–43484,

    [Yanget al., 2024 ] Leshanshui Yang, Clement Chatelain, and Sebastien Adam. Dynamic graph representation learning with neural networks: A survey.Ieee Access, 12:43460–43484,

    2024

  24. [43]

    Towards better dynamic graph learning: New architec- ture and unified library.Advances in Neural Information Processing Systems, 36:67686–67700,

    [Yuet al., 2023 ] Le Yu, Leilei Sun, Bowen Du, and Weifeng Lv. Towards better dynamic graph learning: New architec- ture and unified library.Advances in Neural Information Processing Systems, 36:67686–67700,

    2023

  25. [44]

    Financial knowledge graph based financial report query system.IEEE Access, 9:69766– 69782,

    [Zehraet al., 2021 ] Samreen Zehra, Syed Farhan Mohsin Mohsin, Shaukat Wasi, Syed Imran Jami, Muham- mad Shoaib Siddiqui, and Muhammad Khaliq-Ur- Rahman Raazi Syed. Financial knowledge graph based financial report query system.IEEE Access, 9:69766– 69782,

    2021

  26. [45]

    A dy- namic attributes-driven graph attention network modeling on behavioral finance for stock prediction.ACM Trans- actions on Knowledge Discovery from Data, 18(1):1–29,

    [Zhanget al., 2023 ] Qiuyue Zhang, Yunfeng Zhang, Xunxi- ang Yao, Shilong Li, Caiming Zhang, and Peide Liu. A dy- namic attributes-driven graph attention network modeling on behavioral finance for stock prediction.ACM Trans- actions on Knowledge Discovery from Data, 18(1):1–29,

    2023

  27. [46]

    Graph neural networks: A review of methods and applications.AI open, 1:57–81,

    [Zhouet al., 2020 ] Jie Zhou, Ganqu Cui, Shengding Hu, Zhengyan Zhang, Cheng Yang, Zhiyuan Liu, Lifeng Wang, Changcheng Li, and Maosong Sun. Graph neural networks: A review of methods and applications.AI open, 1:57–81,

    2020

  28. [1967]

    Numerics of gram-schmidt or- thogonalization.Linear Algebra and Its Applications, 197:297–316,

    [Bj¨orck, 1994] ˚Ake Bj ¨orck. Numerics of gram-schmidt or- thogonalization.Linear Algebra and Its Applications, 197:297–316,

    1994

  29. [1991]

    Socialrec: User activ- ity based post weighted dynamic personalized post recom- mendation system in social media

    [Hossainet al., 2024 ] Ismail Hossain, Sai Puppala, Md Ja- hangir Alam, and Sajedul Talukder. Socialrec: User activ- ity based post weighted dynamic personalized post recom- mendation system in social media. InInternational Con- ference on Advances in Social Networks Analysis and Min- ing, pages 263–280. Springer,

    2024

  30. [1994]

    Relgnn: Composite message passing for relational deep learning.arXiv preprint arXiv:2502.06784,

    [Chenet al., 2025 ] Tianlang Chen, Charilaos Kanatsoulis, and Jure Leskovec. Relgnn: Composite message passing for relational deep learning.arXiv preprint arXiv:2502.06784,

    Pith/arXiv arXiv 2025

  31. [2003]

    Classifier ensembles: Select real-world applications.In- formation fusion, 9(1):4–20,

    [Oza and Tumer, 2008] Nikunj C Oza and Kagan Tumer. Classifier ensembles: Select real-world applications.In- formation fusion, 9(1):4–20,

    2008

  32. [2004]

    Asset trees and asset graphs in financial markets

    [Onnelaet al., 2003 ] Jukka-Pekka Onnela, Anirban Chakraborti, Kimmo Kaski, Janos Kertesz, and Antti Kanto. Asset trees and asset graphs in financial markets. Physica Scripta, 2003(T106):48,

    2003

  33. [2008]

    A survey on graph database man- agement techniques for huge unstructured data.Interna- tional Journal of Electrical and Computer Engineering, 8(2):1140,

    [Patilet al., 2018 ] NS Patil, P Kiran, NP Kiran, and Naresh Patel KM. A survey on graph database man- agement techniques for huge unstructured data.Interna- tional Journal of Electrical and Computer Engineering, 8(2):1140,

    2018

  34. [2009]

    Foundations and modeling of dy- namic networks using dynamic graph neural networks: A survey.iEEE Access, 9:79143–79168,

    [Skardinget al., 2021 ] Joakim Skarding, Bogdan Gabrys, and Katarzyna Musial. Foundations and modeling of dy- namic networks using dynamic graph neural networks: A survey.iEEE Access, 9:79143–79168,

    2021

  35. [2012]

    Making a graph database from unstruc- tured text

    [Jeonet al., 2013 ] Seungwoo Jeon, Yohanes Khosiawan, and Bonghee Hong. Making a graph database from unstruc- tured text. In2013 IEEE 16th International Conference on Computational Science and Engineering, pages 981–988. IEEE,

    2013

  36. [2013]

    Lightgbm: A highly efficient gradient boost- ing decision tree.Advances in neural information process- ing systems, 30,

    [Keet al., 2017 ] Guolin Ke, Qi Meng, Thomas Finley, Taifeng Wang, Wei Chen, Weidong Ma, Qiwei Ye, and Tie-Yan Liu. Lightgbm: A highly efficient gradient boost- ing decision tree.Advances in neural information process- ing systems, 30,

    2017

  37. [2014]

    Patient access to medical records and healthcare outcomes: a systematic review.Jour- nal of the American Medical Informatics Association, 21(4):737–741,

    [Davis Giardinaet al., 2014 ] Traber Davis Giardina, Shailaja Menon, Danielle E Parrish, Dean F Sittig, and Hardeep Singh. Patient access to medical records and healthcare outcomes: a systematic review.Jour- nal of the American Medical Informatics Association, 21(4):737–741,

    2014

  38. [2016]

    Gram-schmidt orthogonalization: 100 years and more.Numerical Linear Algebra with Applications, 20(3):492–532,

    [Leonet al., 2013 ] Steven J Leon, ˚Ake Bj ¨orck, and Wal- ter Gander. Gram-schmidt orthogonalization: 100 years and more.Numerical Linear Algebra with Applications, 20(3):492–532,

    2013

  39. [2017]

    Solving linear least squares problems by gram-schmidt orthogonalization.BIT Numer- ical Mathematics, 7(1):1–21,

    [Bj¨orck, 1967] ˚Ake Bj ¨orck. Solving linear least squares problems by gram-schmidt orthogonalization.BIT Numer- ical Mathematics, 7(1):1–21,

    1967

  40. [2018]

    Graph- based semantic web service composition for healthcare data integration.Journal of healthcare engineering, 2017(1):4271273,

    [Arch-Intet al., 2017 ] Ngamnij Arch-Int, Somjit Arch-Int, Suphachoke Sonsilphong, and Paweena Wanchai. Graph- based semantic web service composition for healthcare data integration.Journal of healthcare engineering, 2017(1):4271273,

    2017

  41. [2019]

    Relbench: A benchmark for deep learning on relational databases.Advances in Neural Information Processing Systems, 37:21330–21341,

    [Robinsonet al., 2024 ] Joshua Robinson, Rishabh Ranjan, Weihua Hu, Kexin Huang, Jiaqi Han, Alejandro Dobles, Matthias Fey, Jan Eric Lenssen, Yiwen Yuan, Zecheng Zhang, et al. Relbench: A benchmark for deep learning on relational databases.Advances in Neural Information Processing Systems, 37:21330–21341,

    2024

  42. [2020]

    Transaction logs.Applications of social research methods to questions in information and library science, pages 166–177,

    [Shebleet al., 2009 ] Laura Sheble, Barbara Wildemuth, and Kathy Brennan. Transaction logs.Applications of social research methods to questions in information and library science, pages 166–177,

    2009

  43. [2021]

    Towards structured log analysis

    [Jayathilake, 2012] Dileepa Jayathilake. Towards structured log analysis. In2012 Ninth International Conference on Computer Science and Software Engineering (JCSSE), pages 259–264. IEEE,

    2012

  44. [2022]

    Relational deep learning: Challenges, foundations and next-generation architectures.arXiv preprint arXiv:2506.16654,

    [Dwivediet al., 2025 ] Vijay Prakash Dwivedi, Charilaos Kanatsoulis, Shenyang Huang, and Jure Leskovec. Relational deep learning: Challenges, foundations and next-generation architectures.arXiv preprint arXiv:2506.16654,

    Pith/arXiv arXiv 2025

  45. [2023]

    Exiting the risk as- sessment maze: A meta-survey.ACM Computing Surveys (CSUR), 51(1):1–30,

    [Gritzaliset al., 2018 ] Dimitris Gritzalis, Giulia Iseppi, Alexios Mylonas, and Vasilis Stavrou. Exiting the risk as- sessment maze: A meta-survey.ACM Computing Surveys (CSUR), 51(1):1–30,

    2018

  46. [2024]

    Relational deep learning: Graph representation learning on relational databases.arXiv preprint arXiv:2312.04615,

    [Feyet al., 2023 ] Matthias Fey, Weihua Hu, Kexin Huang, Jan Eric Lenssen, Rishabh Ranjan, Joshua Robinson, Rex Ying, Jiaxuan You, and Jure Leskovec. Relational deep learning: Graph representation learning on relational databases.arXiv preprint arXiv:2312.04615,

    Pith/arXiv arXiv 2023

  47. [2025]

    Empirical evalua- tion of gated recurrent neural networks on sequence mod- eling.arXiv preprint arXiv:1412.3555,

    [Chunget al., 2014 ] Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. Empirical evalua- tion of gated recurrent neural networks on sequence mod- eling.arXiv preprint arXiv:1412.3555,

    Pith/arXiv arXiv 2014