Pith. sign in

REVIEW 4 major objections 6 minor 17 references

A Poisson-Gamma model with an explicit relation-transition kernel predicts sparse temporal links and recovers how relations evolve.

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 · grok-4.5

2026-07-12 06:27 UTC pith:4MXTR6LJ

load-bearing objection Clean multi-relational Poisson–Gamma extension; the relation-transition kernel is the real addition, but the big AUC-PR lifts sit on heavily filtered 500-entity subgraphs. the 4 major comments →

arxiv 2607.02872 v1 pith:4MXTR6LJ submitted 2026-07-03 cs.LG

Poisson-Gamma Modeling of Inter-Relational Dependencies in Dynamic Knowledge Graphs

Nan Fang , Yijun Wang , Hao Liao , Sikun Yang This is my paper
classification cs.LG
keywords dynamic knowledge graphslink predictionPoisson-Gamma modelrelation transitionGamma-Markov processGibbs samplingmulti-relational temporal data
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.

Dynamic knowledge graphs change over time and mix many relation types, yet most predictors treat those relations as independent and struggle when data are sparse. This paper proposes PGRE, a generative model that represents multi-relational temporal links with a Poisson-Bernoulli construction, attaches Gamma latent factors to entities as subjects and objects, and lets shared latent communities mediate cross-relation influence. A Gamma-Markov process equipped with a Dirichlet transition matrix then carries those community weights from one time step to the next, so that the probability of one relation can depend on the history of others. Gibbs sampling with negative-binomial and Chinese-restaurant-table augmentations yields closed-form updates. On filtered ICEWS18, GDELT and WIKI snapshots the model matches or exceeds strong neural and Bayesian baselines on AUC-PR, especially under sparsity, while the learned transition matrix and factor trajectories display self-persistence and structured switches among related relation types.

Core claim

Jointly modeling entity-community factor loadings and an explicit relation-to-relation transition kernel inside a Gamma-Markov process produces competitive or superior link-prediction accuracy on sparse multi-relational temporal knowledge graphs and recovers interpretable patterns of relational evolution.

What carries the argument

PGRE: a Poisson-Bernoulli generative model whose community weights evolve by a Gamma-Markov process whose shape is a linear mixture of all previous relation-specific weights, the mixing coefficients forming a Dirichlet-drawn transition matrix that couples relations across time.

Load-bearing premise

The filtered subgraphs that keep only the most frequent relations and the 500 most active entities still contain the same inter-relational temporal structure the model is meant to learn.

What would settle it

Re-run the identical forecasting protocol on the unfiltered full tensors (or on a random sample of low-frequency relations and low-activity entities) and check whether the AUC-PR gains of PGRE over the strongest baselines disappear or reverse.

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

If this is right

  • Sparse multi-relational forecasting can be improved by an explicit, learnable relation-transition kernel rather than by deeper neural encoders alone.
  • Posterior samples of the transition matrix and community weights supply human-readable maps of how one relation type tends to follow another.
  • The same Gamma-Markov construction can be applied to other sparse multi-type count tensors that evolve over discrete time.
  • Ablation by removing the transition matrix measurably lowers AUC-PR, confirming that cross-relation coupling is a necessary ingredient of the reported gains.

Where Pith is reading between the lines

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

  • If the transition kernel can be made time-varying, the same framework could track regime shifts such as sudden changes in diplomatic or collaboration patterns.
  • The hierarchical Gamma shrinkage already present for communities suggests a natural route to nonparametric inference over the number of active relation types themselves.
  • Because inference is fully Bayesian and closed-form, the model could serve as a calibrated uncertainty source for hybrid systems that combine probabilistic forecasts with large language models.

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 / 6 minor

Summary. The paper proposes PGRE, a Bayesian generative model for dynamic multi-relational knowledge graphs. Multi-relational temporal links are generated via a Bernoulli–Poisson (equivalently Poisson–Bernoulli) construction whose intensity factors into subject/object Gamma entity–community loadings and time- and relation-specific community weights. Those weights evolve under a Gamma–Markov process whose shape is a linear mixture of previous-relation weights through a Dirichlet-prior transition matrix Π, so that inter-relational temporal dependence is modeled explicitly inside shared latent communities. Posterior inference uses Gibbs sampling with negative-binomial–logarithmic / CRT data augmentation and standard Poisson–multinomial and Gamma–Poisson conjugacies (Theorems 1–3, §4 and Supp. §7.5). On moderate-scale filtered versions of ICEWS18, GDELT, and WIKI, PGRE reports competitive or superior AUC-PR (especially on the sparser multi-relational sets), an ablation removing Π (PGE), a relation-cardinality study, runtime comparisons, and visualizations of latent-factor trajectories and transition structure.

Significance. Bayesian modeling of dynamic multi-relational knowledge graphs with an explicit relation-transition kernel is relatively underexplored compared with neural TKG models and single-relation Poisson–Gamma network models. If the empirical advantage holds under more representative evaluation, PGRE would be a useful, interpretable alternative for sparse multi-relational temporal settings, with closed-form conjugacy-based Gibbs updates, public code, an ablation isolating Π, τ-sensitivity checks, and sampling diagnostics. The technical extension from single-relation Gamma–Markov community models to a multi-relational transition kernel is a clear, incremental contribution to the probabilistic-network literature.

major comments (4)
  1. [§5.1, Table 1, Fig. 3] §5.1 preprocessing and Table 1: All reported gains (notably AUC-PR on ICEWS18 0.198 vs. next-best 0.127 and GDELT 0.277 vs. 0.136) are obtained after retaining only the most frequent relations (R=8–15) and then the 500 most active entities. This selection can densify local neighborhoods and reshape the very cross-relation co-occurrence and transition statistics that Π is intended to capture. The R-ablation in Fig. 3 keeps the entity set fixed, so it does not test this assumption. The central claim of advantage on “sparse multi-relational temporal knowledge graphs” therefore remains conditional on a non-standard filtered regime. Please either (i) report results on unfiltered or substantially larger entity sets under the same protocol, or (ii) provide quantitative evidence that the filtered tensors preserve the inter-relational temporal structure of the original TKGs, and state the limitat
  2. [§5.1 Experimental Settings, Table 1] §5.1 evaluation protocol: Performance is reported only as AUC-ROC/AUC-PR on a single future snapshot (tq=T) with binary presence labels. Standard TKG forecasting literature typically reports filtered MRR and Hits@k under negative sampling (or time-aware ranking). Without at least one ranking-based metric on the same splits—or a clear argument why AUC is the right primary metric for this generative formulation—it is hard to place the “competitive or superior” claim relative to the neural baselines the paper cites (Know-Evolve, DyRep, RE-NET). Adding MRR/Hits@k (or an equivalent ranking evaluation) on the same data would make the comparison load-bearing rather than protocol-dependent.
  3. [Supp. §7.1, Table 1, §5.2 PGE ablation] Supp. §7.1 and Table 1 baselines: Single-relation probabilistic models (DPGM, D-NGPPF, G-HSEPM) are trained independently per relation on bipartite undirected conversions, while PGRE jointly models all relations through Π and shared communities. That design difference is legitimate but should be discussed as a confound when attributing AUC-PR gains specifically to the transition kernel rather than to joint multi-relational sharing alone. The PGE ablation helps, but a joint multi-relational baseline without temporal Π (or with a diagonal-only Π) on the same tensor would isolate the contribution more cleanly.
  4. [Supp. §7.3, Figs. 4–5, §5.2] Supp. §7.3 sampling diagnostics and Figs. 4–5: Post-burn-in ESS for some normalized community weights is very low (e.g., 12.99 and 23.63), with strong serial dependence in the trace plots. Interpretability claims that rest on temporal evolution of δ^{(t,r)}_k and on the structure of Π (Figs. 4–5, §5.2) need either longer chains / thinning, multiple-chain diagnostics, or a statement that those figures are qualitative and not fully mixed posterior summaries. Low ESS does not invalidate link-prediction means if predictive functionals mix better, but it weakens the “revealing meaningful relational evolution patterns” claim as currently stated.
minor comments (6)
  1. [Abstract, §1, §5.1] Abstract and §1 claim “particularly in sparse settings,” but §5.1 deliberately keeps the most active entities; clarify what “sparse” refers to (relation density, positive rate, or entity activity) and align wording with the constructed tensors.
  2. [§3.2, Eq. (9)] Eq. (9) Dirichlet prior uses a special diagonal mass ξν_r; the role of ξ versus the off-diagonal ν_r'ν_r construction is easy to miss—add a short sentence on why self-transitions are parameterized separately.
  3. [Fig. 1, §3.2] Fig. 1 caption uses y_uv,t while the model uses M^{(t,r)}_{ij} / x^{(t,r)}_{ij}; unify notation between the motivational figure and the generative model.
  4. [Table 2, §5.1] Table 2: PGRE is fastest on WIKI but slower than several count models on GDELT; a one-sentence complexity discussion (dependence on R, K, T, and nonzero events) would help readers interpret the runtime trade-off.
  5. [§4, Supp. Algorithm 1] Supp. Algorithm 1 and the main-text sampling order for ρ^{(t)}_k / forward CRT variables are slightly hard to follow; a short forward–backward sketch in the main §4 would improve readability without repeating the full supplement.
  6. [§2] Related work could briefly position against other multi-relational Bayesian TKG / tensor models beyond the count-reshaping line (PGDS/PRGDS/NBRGDS), even if only to note differences in how relation interactions are encoded.

Circularity Check

0 steps flagged

No load-bearing circularity: generative Poisson–Gamma model with conjugacy-based Gibbs inference is self-contained and evaluated by held-out link prediction; minor self-citations appear only as baselines.

full rationale

The paper defines a fully generative hierarchical model (Eqs. 4–10) whose parameters (entity–community factors ϕ, ψ; community weights δ under a Gamma–Markov process with Dirichlet transition kernel Π; hyperparameters) are given independent priors and updated by standard Gamma–Poisson / Dirichlet–multinomial conjugacy plus CRT augmentation (Theorems 1–3, cited to Zhou 2015). Link probabilities are obtained from the posterior predictive under the Bernoulli–Poisson link; evaluation is ordinary temporal hold-out (train t=1…T-2, test t=T) measured by AUC-PR/ROC. Nothing is defined in terms of the quantity being predicted, no parameter fitted on a subset is re-labeled a “prediction,” and no uniqueness theorem or ansatz is imported from the authors’ prior work to force the present construction. Self-citations (Yang & Koeppl 2018, Acharya et al. 2015, Yu et al. 2025) appear only in Related Work and as empirical baselines; they are not used to justify the generative assumptions or the inference algorithm. The derivation chain therefore does not reduce to its own inputs by construction. Score 1 reflects only the presence of ordinary author-overlap citations that are not load-bearing.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 2 invented entities

The model rests on standard conjugate Bayesian network assumptions plus a small set of free hyper-parameters and one new structural object (the relation-transition matrix). No physical constants or unproved mathematical lemmas are required; the load-bearing modeling choices are the Gamma-Markov evolution, the Dirichlet prior on Π, and the Bernoulli-Poisson link.

free parameters (4)
  • K (number of latent communities) = 200
    Fixed at 200 for all experiments; the infinite-limit shrinkage argument is stated but not used in practice.
  • τ (temporal scale of Gamma-Markov) = 1.0
    Default 1.0; sensitivity table shows modest variation but still a free scale.
  • a0, a1 (Gamma shape for entity-community factors)
    Hyper-parameters drawn from further Gammas or set by convention; control sparsity of memberships.
  • γ0, β, ξ and related Dirichlet/Gamma hyper-parameters
    Control concentration of the relation-transition prior and initial community weights; set by standard weakly-informative choices.
axioms (4)
  • domain assumption Bernoulli-Poisson link: presence of a multi-relational edge is the indicator that a Poisson count with rate sum_k δ_k ϕ_i ψ_j is at least 1.
    Standard sparse-graph modeling device (Zhou et al.); invoked in Eqs. (1)-(3).
  • ad hoc to paper Community weights under each relation evolve as a Gamma Markov process whose shape is a linear mixture of previous-relation weights via a stochastic matrix Π.
    Core modeling choice that extends single-relation Gamma-Markov models; stated in Eqs. (7)-(9).
  • domain assumption Entity-community factor loadings are independent Gamma random variables (subject and object roles separate).
    Inherited from earlier Poisson-Gamma network models; Eqs. (6).
  • standard math Dirichlet prior on each row of the relation-transition matrix with concentration modulated by ν_r.
    Conjugate prior for multinomial transition counts; Eq. (9).
invented entities (2)
  • Relation-transition matrix Π inside the Gamma-Markov community process no independent evidence
    purpose: Couples the temporal evolution of the same latent community across different relation types, enabling inter-relational dependence.
    The central structural novelty; learned from data and visualized in Fig. 5. No independent physical existence claimed outside the model.
  • Time- and relation-specific community weights δ^{(t,r)}_k no independent evidence
    purpose: Mediate how strongly each latent community contributes to each relation at each time.
    Standard latent-factor construction specialized to the multi-relational temporal setting.

pith-pipeline@v1.1.0-grok45 · 21885 in / 2979 out tokens · 30707 ms · 2026-07-12T06:27:38.658301+00:00 · methodology

0 comments
read the original abstract

Dynamic knowledge graphs are ubiquitous in today's AI applications, as we represent molecular structures, social relationships, and language information using these graph models. As knowledge graphs evolve over time and are often noisy and incomplete, modeling their temporal and relational dependencies becomes crucial for downstream tasks. To address these challenges, this paper proposes PGRE (Poisson-Gamma Relational Evolution), a probabilistic model for modeling inter-relational dependencies in dynamic knowledge graphs. PGRE represents multi-relational temporal links via a Poisson-Bernoulli formulation. It introduces Gamma-distributed latent variables to capture entity-factor associations and cross-relation dependencies mediated by shared latent communities. A Gamma Markov process further models the temporal evolution of these latent variables, enabling principled characterization of relational dynamics. Experiments on benchmark datasets show that PGRE achieves competitive performance in link prediction, particularly in sparse settings, while revealing meaningful relational evolution patterns in dynamic knowledge graphs.

Figures

Figures reproduced from arXiv: 2607.02872 by Hao Liao, Nan Fang, Sikun Yang, Yijun Wang.

Figure 1
Figure 1. Figure 1: Illustration of a relation transition prior and its [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The overall structure of the proposed PGRE model. The upper part illustrates the hierarchical Gamma–Dirichlet [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of AUC-ROC and AUC-PR under different relation counts [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Temporal evolution of latent factors for six representative relations: [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Transition probabilities among the 25 most ac [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Trace plots of selected representative relation transition parameters over 2,000 Gibbs sampling iterations. The [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Trace plots of selected normalized temporal community weights at [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Full probabilistic graphical model of PGRE. [PITH_FULL_IMAGE:figures/full_fig_p020_8.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

17 extracted references · 7 linked inside Pith

  1. [1]

    G2s: A general-to-specific learning framework for temporal knowledge graph fore- casting with large language models

    Long Bai, Zixuan Li, Xiaolong Jin, Jiafeng Guo, Xueqi Cheng, and Tat-Seng Chua. G2s: A general-to-specific learning framework for temporal knowledge graph fore- casting with large language models. InFindings of the Association for Computational Linguistics: ACL 2025, pages 20927–20938,

    2025

  2. [2]

    Hyte: Hyperplane-based temporally aware knowledge graph embedding

    Shib Sankar Dasgupta, Swayambhu Nath Ray, and Partha Talukdar. Hyte: Hyperplane-based temporally aware knowledge graph embedding. InProceedings of the 2018 conference on empirical methods in natural language processing, pages 2001–2011,

    2018

  3. [3]

    Learning sequence encoders for temporal knowl- edge graph completion.arXiv preprint arXiv:1809.03202,

    Alberto García-Durán, Sebastijan Dumanˇci´c, and Mathias Niepert. Learning sequence encoders for temporal knowl- edge graph completion.arXiv preprint arXiv:1809.03202,

    Pith/arXiv arXiv

  4. [4]

    Negative- binomial randomized gamma markov processes for het- erogeneous overdispersed count time series.arXiv preprint arXiv:2402.18995,

    Rui Huang, Sikun Yang, and Heinz Koeppl. Negative- binomial randomized gamma markov processes for het- erogeneous overdispersed count time series.arXiv preprint arXiv:2402.18995,

    Pith/arXiv arXiv

  5. [5]

    Temporal graph benchmark for ma- chine learning on temporal graphs.Advances in Neural Information Processing Systems, 36:2056–2073,

    Shenyang Huang, Farimah Poursafaei, Jacob Danovitch, Matthias Fey, Weihua Hu, Emanuele Rossi, Jure Leskovec, Michael Bronstein, Guillaume Rabusseau, and Reihaneh Rabbany. Temporal graph benchmark for ma- chine learning on temporal graphs.Advances in Neural Information Processing Systems, 36:2056–2073,

    2056

  6. [6]

    Re- current event network: Autoregressive structure infer- ence over temporal knowledge graphs.arXiv preprint arXiv:1904.05530,

    Woojeong Jin, Meng Qu, Xisen Jin, and Xiang Ren. Re- current event network: Autoregressive structure infer- ence over temporal knowledge graphs.arXiv preprint arXiv:1904.05530,

    Pith/arXiv arXiv 1904

  7. [7]

    Deriving validity time in knowledge graph

    Julien Leblay and Melisachew Wudage Chekol. Deriving validity time in knowledge graph. InCompanion pro- ceedings of the web conference 2018, pages 1771–1776,

    2018

  8. [8]

    Kalev Leetaru and Philip A. Schrodt. Gdelt: Global data on events, location, and tone, 1979–2012. InProceedings of the International Studies Association Annual Convention, pages 1–49. Citeseer,

    1979

  9. [9]

    Dy- namic anticipation and completion for multi-hop rea- soning over sparse knowledge graph.arXiv preprint arXiv:2010.01899,

    Xin Lv, Xu Han, Lei Hou, Juanzi Li, Zhiyuan Liu, Wei Zhang, Yichi Zhang, Hao Kong, and Suhui Wu. Dy- namic anticipation and completion for multi-hop rea- soning over sparse knowledge graph.arXiv preprint arXiv:2010.01899,

    Pith/arXiv arXiv 2010

  10. [10]

    Tempo- ral graph networks for deep learning on dynamic graphs

    Emanuele Rossi, Ben Chamberlain, Fabrizio Frasca, Davide Eynard, Federico Monti, and Michael Bronstein. Tempo- ral graph networks for deep learning on dynamic graphs. arXiv preprint arXiv:2006.10637,

    Pith/arXiv arXiv 2006

  11. [11]

    Kgat: Knowledge graph attention network for recommendation

    Xiang Wang, Xiangnan He, Yixin Cao, Meng Liu, and Tat- Seng Chua. Kgat: Knowledge graph attention network for recommendation. InProceedings of the 25th ACM SIGKDD international conference on knowledge discov- ery & data mining, pages 950–958, 2019a. Xiang Wang, Dingxian Wang, Canran Xu, Xiangnan He, Yixin Cao, and Tat-Seng Chua. Explainable reasoning over...

    2021

  12. [12]

    Estimating latent popu- lation flows from aggregated data via inversing multi- marginal optimal transport

    Sikun Yang and Hongyuan Zha. Estimating latent popu- lation flows from aggregated data via inversing multi- marginal optimal transport. InProceedings of the 2023 SIAM International Conference on Data Mining, SDM, pages 181–189,

    2023

  13. [13]

    Qa-gnn: Reasoning with lan- guage models and knowledge graphs for question answer- ing.arXiv preprint arXiv:2104.06378,

    Michihiro Yasunaga, Hongyu Ren, Antoine Bosselut, Percy Liang, and Jure Leskovec. Qa-gnn: Reasoning with lan- guage models and knowledge graphs for question answer- ing.arXiv preprint arXiv:2104.06378,

    Pith/arXiv arXiv

  14. [14]

    Tgl: A general frame- work for temporal gnn training on billion-scale graphs

    Hongkuan Zhou, Da Zheng, Israt Nisa, Vasileios Ioannidis, Xiang Song, and George Karypis. Tgl: A general frame- work for temporal gnn training on billion-scale graphs. arXiv preprint arXiv:2203.14883,

    Pith/arXiv arXiv

  15. [15]

    Probabilistic Models for Dynamic Networks.DPGM, D-NGPPF, and G-HSEPM are designed for dynamic single- relation networks

    Poisson–Gamma Modeling of Inter-Relational Dependencies in Dynamic Knowledge Graphs (Supplementary Material) Nan Fang1,2 Yijun Wang1,3 Hao Liao2 Sikun Yang*1,4 1School of Computing and Information Technology, Great Bay University, Dongguan, Guangdong, China 2College of Computer Science and Software Engineering, Shenzhen University, Shenzhen, Guangdong, Ch...

    2000

  16. [16]

    Unless otherwise specified, the learning rate and batch size are set to 0.0001 and 200, respectively

    The model checkpoint achieving the best validation performance is selected for testing. Unless otherwise specified, the learning rate and batch size are set to 0.0001 and 200, respectively. To reduce the effect of randomness, each experiment was repeated five times with different random seeds, and the reported results correspond to average performance. Al...

    arXiv 2067

  17. [17]

    The first 1,000 iterations are treated as burn-in, and the remaining 1,000 iterations are retained for posterior estimation

    Figure 6: Trace plots of selected representative relation transition parameters over 2,000 Gibbs sampling iterations. The first 1,000 iterations are treated as burn-in, and the remaining 1,000 iterations are retained for posterior estimation. The trajectories exhibit an initial adaptation period followed by comparatively more stable post-burn-in behavior....

    1991