Pith. sign in

REVIEW 1 major objections 5 minor 32 references

Extended feature allocation models

T0 review · 1 major / 5 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 claims that, within extended feature allocation models, the predictive distribution for new features depends only on sample size exactly when the prior point process is Poisson, and only on sample size plus the number of…

desk verdict The Palm-calculus framework and the DPP prior are genuinely useful, but the sufficientness theorems are over-stated: the 'only if' directions rest on an unproved identifiability step and, as written, appear to be false. read the letter →

arxiv 2502.10257 v3 pith:4IRJ3MIE submitted 2025-02-14 math.ST stat.MEstat.TH

Mario Beraha , Federico Camerlenghi , Lorenzo Ghilotti This is my paper · ORCID
classification math.STstat.MEstat.TH MSC 60G5562F1560G57
keywords featureallocationmodelsIndianbuffetprocessBayesiannonparametricssufficientnesspostulatesPalmcalculuspointprocessesdeterminantalpredictivedistributions
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

Feature allocation models describe data in which each observation carries several labels, and standard versions assume feature labels are independent. This paper develops a unified Bayesian treatment of extended feature allocation models, where feature labels and their probabilities are generated by an arbitrary point process, so labels may attract or repel. Using Palm calculus, it derives closed-form marginal, posterior, and predictive distributions and then proves two sufficientness postulates: new-feature predictions depend solely on sample size if and only if the prior point process is Poisson, and depend only on sample size plus the observed feature count if and only if the prior is mixed Poisson or mixed binomial. These results unify earlier, example-specific characterizations for completely random measures, stable beta scaled processes, and product-form feature models, and they give practitioners a criterion for choosing priors in the unseen-feature problem. The paper also introduces a determinantal-point-process prior whose predictions depend on observed feature locations and applies it to estimating forest size and locating unseen trees.

What carries the argument

The load-bearing object is the reduced Palm kernel of the prior point process $\Psi = \sum_j \delta_{(X_j,S_j)}$ on $X \times (0,1]$. For conditioning points $(x,s)$, the reduced Palm version $\Psi^!_{x,s}$ is the law of the remaining atoms after forcing atoms at $x$ with marks $s$ and removing them; it enters Theorem 2's posterior representation and, through thinning, Theorem 3's predictive law for $Z'_{n+1}$. Lemma 1 and Lemma 2 are the identities that carry the argument: the reduced Palm law is location-independent if and only if $\Psi$ is Poisson, and it depends only on the number of conditioning points if and only if $\Psi$ is mixed Poisson or mixed binomial. Everything else in the Bayesian analysis is a Palm-calculus computation: the factorial moment disintegration, the Campbell-Little-Mecke formula, and the Laplace functional ratio that defines the predictive law.

What would settle it

On a finite label space, enumerate possible reduced Palm kernels for a two-point prior and compute the induced predictive law, the Bernoulli-process thinning in equation (7), for both a Poisson and a non-Poisson kernel with the same factorial moments; if two different kernels yield the same predictive law for every sample, the only-if direction of Theorem 4 fails. Even a single analytic pair of non-Poisson processes with identical new-feature predictive laws would refute the stated characterization.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is a pair of characterization theorems. Theorem 4 states that, within model (3), the distribution of the new-feature process $Z'_{n+1}$ from Theorem 3 is a function of the sample size $n$ alone if and only if the underlying point process $\Psi$ is Poisson. Theorem 5 states that this distribution is a function of $n$ and the number of distinct features $k$ alone if and only if $\Psi$ is a mixed Poisson or mixed binomial process. The proof route that supplies these results is a characterization of the reduced Palm kernel: a point process whose reduced Palm versions $\Psi^!_{x,s}$ have a law independent of the conditioning points must be Poisson, and a process whose reduced Palm laws depend only on the number of conditioning points must be mixed Poisson or mixed binomial. Thus the simplicity of predictive distributions is exactly the simplicity of Palm kernels, and the paper derives a characterization of the Poisson process as a byproduct.

Load-bearing premise

The only-if directions of both theorems assume that the distribution of the thinned new-feature process $Z'_{n+1}$ uniquely determines the reduced Palm kernel of $\Psi$; the proof states that this is clear rather than proving that the map from Palm kernels to predictive laws is injective.

Editorial extensions

If this is right

  • Under a completely random measure (Poisson) prior, no predictive statement about new features can use the observed labels or the frequency spectrum; the entire new-feature law is fixed by the sample size.
  • To obtain predictions that depend on both sample size and number of distinct features, it is necessary and sufficient to use a mixed Poisson or mixed binomial prior, and stable beta scaled processes and product-form feature models are recovered as special cases.
  • Any extended feature prior outside these classes yields new-feature predictions that depend on labels or frequencies, giving a principled reason to choose such priors for the unseen-feature problem.
  • The reduced-Palm characterization is a standalone result: a point process whose reduced Palm law does not depend on the conditioning point is Poisson.
  • The independently marked determinantal point process prior makes the predictive locations of new features depend on the observed locations, which is what allows the forest application to locate unseen trees.

Reading between the lines

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

  • If the identifiability step that the supplementary proof marks as clear fails, the only-if directions of Theorems 4 and 5 would collapse to one-way statements; the theorems should then be reformulated as characterizing the simplest predictive laws rather than the priors themselves.
  • A natural cross-check is to test the forest application against a Poisson prior with matching mean measure: the repulsive prior should concentrate posterior predictive mass closer to the observed configuration, and a formal calibration of coverage over repeated surveys would quantify the gain.
  • The sufficientness postulates constrain only the new-feature component $Z'_{n+1}$; they say nothing about the probability of re-observing old features, so prior selection for joint prediction of old and new features would need additional predictive functionals.
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.

Referee Report

1 major / 5 minor

Summary. This paper introduces extended feature allocation models, in which the latent feature process Ψ is an arbitrary simple point process on X×(0,1] and observations are i.i.d. Bernoulli processes directed by the random measure μ(B)=∫ s 1_B(x)Ψ(dx ds). The main theoretical contributions are: (i) a general Palm-calculus treatment giving the marginal, posterior, and predictive laws (Theorems 1-3); (ii) two sufficientness postulates — the new-feature predictive law depends on the sample only through n iff Ψ is Poisson (Theorem 4), and only through n and k iff Ψ is mixed Poisson or mixed binomial (Theorem 5); and (iii) specialization to Poisson, mixed Poisson, mixed binomial, and independently marked determinantal point process priors, with an application to forest surveys.

Significance. Should the characterizations be fully established, Theorems 4 and 5 are substantial: they unify previous sufficientness results for CRMs, scaled stable beta processes, and product-form EFPMs within a single point-process framework, and they give practitioners a clear criterion for prior selection. The Palm-calculus machinery in Theorems 1-3 is clean, and the specializations to Poisson, mixed Poisson, and mixed binomial priors are consistent with known results. The paper also offers a novel characterization of the Poisson process via its reduced Palm kernel and a useful DPP-based spatial model. However, the only-if directions of the two central theorems rest on an unproved identifiability step, so the main characterization claim is not yet fully demonstrated.

major comments (1)
  1. [Supplementary S5.1 (proofs of Theorems 4 and 5)] The only-if directions of Theorems 4 and 5 are not proved as written. The proof says "it is clear that the predictive distribution of Z'_{n+1} depends on the sampling information as the law of μ', or equivalently Ψ', in Theorem 2 does," but this hides a nontrivial inversion. The observable object Z'_{n+1} is a Bernoulli process directed by μ', and μ' is a mixture over S* of tilted reduced Palm versions of Ψ (Theorem 2(i)-(ii) and Eq. (7)). To conclude from constancy of the predictive law in (x*,m) that each reduced Palm kernel Ψ!_{x*,s*} is independent of (x*,s*), one needs three inversions: (a) the law of a Cox process determines the law of its directing random measure; (b) the law of μ' determines the law of the marked point process Ψ'; and (c) the mixture over S* with weights proportional to s*^m(1-s*)^{n-m}ρ^{(k)}(ds*|x*) E[e^{∫ n log(1-t)Ψ!_{x*,s*}}] determines the family of reduced Palm kernels. Part (a) is standard but not stated; part (c) is nontrivial and no argument is provided. Without (c), different reduced Palm kernels could in principle produce the same mixture predictive law for all observed x*,m, so Lemma 1 cannot be invoked. The same gap affects Theorem 5, whose proof says only "arguing as in the proof of Theorem 4." This is a load-bearing point for the paper's central characterization result.
minor comments (5)
  1. [Section 3.1, Theorem 1] The displayed formula is a density with respect to \tilde m^{(k)}_ξ(dx*), not a probability itself; the wording "The probability ... is" should be "the probability measure has density" or similar.
  2. [Section 7, last paragraph] The claim that "for any Poisson, mixed Poisson, or mixed binomial prior, the probability of re-observing a feature depends exclusively on the sample size and the frequency of that feature" is too strong when the kernel ρ(ds|x) depends on the label x; in Corollaries 1-3 the marginal density of S*_ℓ is proportional to s^{m_ℓ}(1-s)^{n-m_ℓ}ρ(ds|x*_ℓ), so it depends on x*_ℓ unless ρ is label-independent.
  3. [Supplementary S5.2, proof of Lemma 1] The last step of the proof is terse: after showing that each Φ!_{x} is Poisson, the conclusion that Φ itself is Poisson invokes the fact that the family of Palm distributions, together with the mean measure, characterizes the law of a simple point process; this should be stated explicitly with a precise reference.
  4. [Section 6.1] The sentence "the infinitesimal probability that an unobserved tree would occupy position dx equals E{Ψ′(dx × (0,1])}" uses "probability" where the object is a mean measure; please rephrase to avoid confusion with a density.
  5. [Section 2.3] The claim that "extended feature allocation models induce the entire class of regular feature allocation models admitting an efpf" is nontrivial and is stated without proof; a precise reference or a short argument should be supplied.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular reduction: Theorems 4–5 are proved from independent Palm-characterization lemmas; self-citations are contextual and the empirical Bayes fit does not enter the theoretical claims.

full rationale

The central characterization theorems are not circular. Theorem 4 is reduced in Supplementary Section S5.1 to Lemma 1, and Theorem 5 to Lemma 2; both lemmas are stated independently of the predictive setup and proved from classical point-process facts, with Lemma 2 built on Kallenberg (1973) and Lemma 1 on Slivnyak–Mecke and the Palm characterization of point-process laws. The 'if' directions are independently verified by the concrete posterior and predictive computations in Corollaries 1–3, which derive the Poisson, mixed Poisson, and mixed binomial predictive forms from Theorem 2 rather than from the sufficientness postulates. Thus the statements do not assume their conclusions. The paper's use of earlier work by Camerlenghi et al. (2023) and Ghilotti et al. (2024) is for context and for reconciling scaled processes with Theorem 5; it is not the proof of Theorems 4–5. The empirical Bayes selection of DPP hyperparameters in Section 6.1 is standard estimation practice and is confined to the application; it does not feed back into the theoretical derivations. One genuine gap should be flagged, although it is not circularity: Supplementary Section S5.1 asserts 'it is clear' that the predictive law of Z'_{n+1} depends on the sample exactly as the law of mu' or Psi' in Theorem 2 does, and the subsequent proof transfers the sufficientness assumption to a property of the reduced Palm kernel. The needed identifiability step—that the mixture over S* and the Bernoulli thinning can be inverted to recover the Palm-kernel family—is not demonstrated. If false, the 'only if' directions would be incomplete; however, this is an omitted identifiability argument rather than a reduction of the conclusion to its own input, so it does not affect the circularity score.

Assumptions & free parameters 2 free parameters · 6 assumptions · 0 invented entities

The central theoretical results rest on standard point process theory and domain-specific exchangeability assumptions. The key unstated assumption is the identifiability of the reduced Palm kernel from the predictive law of the Bernoulli process, which is load-bearing for the sufficientness postulates. The empirical Bayes DPP hyperparameters are free parameters fitted to the data used for evaluation. No new physical or mathematical entities are postulated beyond the modeling framework itself.

free parameters (2)
  • DPP hyperparameters (a, b, rho, alpha) = Estimated by empirical Bayes, values not reported
    Section 6.1 maximizes the marginal likelihood over the beta mark parameters and the Gaussian DPP parameters; these estimated values drive all figures and the spruces predictions.
  • Simulation constants (marks Beta(1,5) or Beta(1,20), rho=100, alpha=0.0535) = Hand-chosen
    These are data-generating or semi-synthetic settings chosen by the authors for the experiments, not fitted to data, but they influence the reported results.
assumptions (6)
  • standard math Palm calculus and Campbell-Little-Mecke formulas for locally finite point processes
    Used throughout Sections 3 and S4 to derive marginal, posterior, and predictive laws.
  • standard math Kallenberg (1973) Theorem 5.3 characterization of mixed Poisson/binomial processes via Palm kernels
    Basis for Lemma 2 and Theorem 5; the paper verifies it extends to locally finite (infinite activity) processes.
  • domain assumption The statistical model (3): Zi | mu iid BeP(mu), mu is a functional of a simple point process Psi, with conditionally i.i.d. Bernoulli indicators A_ij with probabilities S_j > 0
    Definition of extended feature allocation model in Section 2.2.
  • domain assumption The k-th factorial moment measures of Psi are sigma-finite for all k, and Z_i(X) < infinity a.s.
    Assumed in Section 3 and Remark 1; needed for existence of Palm distributions and for the Poisson/mixed Poisson treatment.
  • ad hoc to paper Identifiability: the law of Z'_n+1 determines the law of mu' and hence the reduced Palm kernel
    Unstated but needed for the proofs of Theorems 4 and 5; see supplementary S5.1.
  • ad hoc to paper In the forest application, tree locations follow a Gaussian determinantal point process and detection probabilities are i.i.d. beta
    Model choice in Section 6.1, not derived from data.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Extended feature allocation models." pith.science (2026). https://pith.science/paper/4IRJ3MIE

@misc{pith2026250210257,
  author       = {Pith},
  title        = {Pith review of: Extended feature allocation models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4IRJ3MIE}},
  note         = {Machine review of arXiv:2502.10257}
}
read the original abstract

Feature allocation models are Bayesian nonparametric tools tailored to data in which each observation can simultaneously exhibit multiple characteristics, or features. A fundamental limitation of standard formulations is that feature labels are assumed to be independent and identically distributed, and therefore play no role in posterior inference. The present paper introduces a unified Bayesian framework for extended feature allocation models, in which feature labels and proportions are modeled jointly, thereby enabling the simultaneous discovery of features and learning of dependencies among their labels. Building on point process theory, we develop a full Bayesian analysis of these models. Within this general setting, we also characterize previously proposed priors as those leading to poor predictive distributions, which cannot capture label dependencies and are insensitive to the observed frequency spectrum. Our methodology is designed to move beyond such standard formulations by leveraging the information carried by feature labels. We demonstrate the usefulness of our approach by introducing: (i) a Cox process prior that clusters genomic variant embeddings while predicting new variants and new variant clusters; (ii) a determinantal point process prior for repeated forest surveys, where prediction concerns both the number and the locations of unobserved trees.

Figures

Figures reproduced from arXiv: 2502.10257 by the authors.

Figure 1
Figure 1. Posterior distribution of the total number of trees in the synthetic scenario of Sec [PITH_FULL_IMAGE:figures/full_fig_p027_1.png] view at source ↗
Figure 2
Figure 2. Locating the unobserved trees for n = 15 in the synthetic scenario of Section 6.2: infinitesimal probability of observing an unseen tree in a given location. Left plot: the mean measure of ξ ′ . Right plot: the mean measure of ξ ! x∗ . The red dots represent the observed trees in the sample. The black crosses indicate the unseen trees. Note that the color scales of the two plots are different. obtained by estimating… view at source ↗
Figure 3
Figure 3. Locating the unobserved trees for n ∈ {10, 20, 30} in the analysis of the spruces dataset of Section 6.3: infinitesimal probability of observing an unseen tree in a given location. The three plots report Mξ ′ for the three sample sizes. The red dots represent the observed trees in the sample. The black crosses indicate the unseen trees. Note that the plots have different color scales. We analyze the spruces dataset … view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

32 extracted references · 30 canonical work pages

  1. [1]

    Franzolini, A

    Ascolani, F., B. Franzolini, A. Lijoi, and I. Prünster (2024). Nonparametric priors with full- range borrowing of information.Biometrika 111(3), 945–969. Ayed, F. and F. Caron (2021). Nonnegative bayesian nonparametric factor models with com- pletely random measures.Statistics and Computing 31, 1–24. Bacallado, S., M. Battiston, S. Favaro, and L. Trippa (...

    work page 2024

  2. [2]

    This distinction, while subtle, is crucial

    However, it is important to note that in Kallenberg’s theorem, a point process is defined as a random boundedly finite counting measure, whereas we consider a point process as a random locally finite counting measure. This distinction, while subtle, is crucial. Working with boundedly finite measures excludes Poisson processes with infinite activity. Speci...

    work page 1973

  3. [4]

    The marginal distribution ofZ is recovered from Theorem 1 as follows. SinceΨ is an independently marked process with ground processξ and mark kernel H, then from (Baccelli et al., 2020, Proposition 3.2.14), the reduced Palm versionΨ! x,s is still an independently marked process, with ground process ξ! x and mark kernel H, thus it does not depend on s. By ...

    work page 2020

  4. [7]

    Lavancier, F., J

    Cambridge University Press. Lavancier, F., J. Møller, and E. Rubak (2015). Determinantal point process models and sta- tistical inference. Journal of the Royal Statistical Society: Series B (Statistical Methodol- ogy) 77(4), 853–877. Lavancier, F. and E. Rubak (2023, October). On simulation of continuous determinantal point processes. Statistics and Compu...

    work page 2015

  5. [12]

    Kallenberg, O

    Springer. Kallenberg, O. (2021).Foundations of Modern Probability. Springer. Last, G. and M. Penrose (2017). Lectures on the Poisson process, Volume

    work page 2021

  6. [18]

    Ghilotti, L., F

    MIT Press. Ghilotti, L., F. Camerlenghi, and T. Rigon (2024). Bayesian analysis of product feature allo- cation models. arXiv preprint arXiv:2408.15806. Gnedin, A. and J. Pitman (2005). Exchangeable Gibbs partitions and Stirling triangles.Zap. Nauchn. Sem. S.-Peterburg. Otdel. Mat. Inst. Steklov. (POMI) 325, 83–102. 34 Good, I.J.andG.H.Toulmin(1956). Then...

    arXiv 2024

  7. [19]

    Thibaux, R. and M. I. Jordan (2007). Hierarchical beta processes and the indian buffet process. In AISTATS, Volume 2, pp. 564–571. Titsias, M. (2007). The infinite gamma-poisson feature model. In J. Platt, D. Koller, Y. Singer, and S. Roweis (Eds.),Advances in Neural Information Processing Systems, Volume

    work page 2007

  8. [20]

    Sufficientness

    Cur- ran Associates, Inc. 38 Xie, F. and Y. Xu (2019). Bayesian repulsive gaussian mixture model.Journal of the American Statistical Association, 187–203. Zabell, S. L. (1982). W. E. Johnson’s "Sufficientness" Postulate.The Annals of Statistics 10(4), 1090 –

    work page 2019

Show all 32 references
  1. [22]

    Møller, J

    Curran Associates, Inc. Møller, J. and R. P. Waagepetersen (2003).Statistical inference and simulation for spatial point processes. CRC Press. Muliere, P. and S. Walker (1999). A characterization of a neutral to the right prior via an extension of Johnson’s sufficientness post...

    2003

  2. [23]

    is equivalent to R X×(0,1] sν(dx ds) < ∞ (Proposition S2)

    Indeed, when Ψ is a Poisson process with mean measureν, the condition Zi(X) < ∞ a.s. is equivalent to R X×(0,1] sν(dx ds) < ∞ (Proposition S2). Therefore, for anyϵ >0, ν(X × (ϵ, 1]) = Z X×(ϵ,1] ν(dx ds) ≤ ϵ−1 Z X×(ϵ,1] sν(dx ds) < ∞. Thus, it follows thatν is σ-finite, as well...

    2021

  3. [24]

    Let LΦ be the Laplace functional of Φ. By (Baccelli et al., 2020, Proposition 3.2.1), for any measurable functions f, g: X → R+, the Palm distribution of a point process satisfies ∂ ∂t LΦ(f + tg) t=0 = − Z X g(x)LΦx(f )MΦ(dx). Since Φ is a mixed binomial processMB(ν, qM ), its...

    2020

  4. [25]

    Pitman, J

    Curran Associates, Inc. Pitman, J. (1995). Exchangeable and partially exchangeable random partitions. Probability theory and related fields 102(2), 145–158. Pitman, J. (1996). Some developments of the Blackwell-MacQueen urn scheme. InStatistics, probability and game theory, Vo...

    1995

  5. [26]

    Hence, Φx = δx + Φ! x where Φ! x is distributed as in the statement

    E(M ) E e−f (X) k # MΦ(dx). Hence, Φx = δx + Φ! x where Φ! x is distributed as in the statement. The proof for the Palm distribution of orderk follows by induction, by using the Palm algebra property (Baccelli et al., 2020, Proposition 3.3.9), i.e.,Φ! (x1,x2) d = Φ! x1 ! x2 , ...

    2020

  6. [27]

    Lemma S5

    Secondly, we state another key result which we need for the proof, corresponding to (Kallen- berg, 2017, Theorem 3.7). Lemma S5. Let Φ be a point process onX and let Cj ↑ X, Cj ∈ X and relatively compact, j ≥

    2017

  7. [28]

    S4 Proof of Theorems 1 and 2 The proof is based on the study of the Laplace functional of µ

    Finally, by an application of Lemma S5, we conclude thatΦ is either a mixed Poisson or mixed binomial process. S4 Proof of Theorems 1 and 2 The proof is based on the study of the Laplace functional of µ. We remind that, for any measurable function f : X → R+, the Laplace funct...

    2017

  8. [29]

    S5.2 Proof of Lemma 1 If Φ is a Poisson process, thenΦ! x d = Φ ! y d = Φ by the multivariate Mecke equation (Last and Penrose, 2017, Theorem 4.4)

    Then, the proof follows from Lemma 1 which characterizes the Poisson process as the unique point process for which the reduced Palm kernelΨ! x∗,s∗ does not depend on(x∗, s∗). S5.2 Proof of Lemma 1 If Φ is a Poisson process, thenΦ! x d = Φ ! y d = Φ by the multivariate Mecke eq...

    2017

  9. [30]

    Using this characterization, proceed as follows to prove Lemma 2

    S5.4 Proof of Lemma 2 The proof of our lemma requires the use of Lemma S3. Using this characterization, proceed as follows to prove Lemma 2 . First, (ii) implies (i) by using (Kallenberg, 1973, Theorem 5.3). Indeed, by settingk = 1, condition (ii) states thatΦ! x does not depe...

    1973

  10. [31]

    Moreover, since Ψ is a Poisson process, by Lemma 1, it holdsΨ! x∗,s d = Ψ

    In particular, we first observe that˜m(k) ξ (dx∗) = Qk ℓ=1 G0(dx∗ ℓ ) and ρ(k)(ds | x∗) = Qk ℓ=1 ρ(dsℓ | x∗ ℓ ). Moreover, since Ψ is a Poisson process, by Lemma 1, it holdsΨ! x∗,s d = Ψ. Consequently, the expected value in the marginal expression of Theorem 1 equals E n e R X...

    2017

  11. [32]

    , S∗ k) and the law ofµ′, conditionally to S∗

    For the posterior distribution ofµ, expressed in Theorem 2, we need to determine the law of the vectorS∗ = (S∗ 1 , . . . , S∗ k) and the law ofµ′, conditionally to S∗. From point (i) of Theorem 2 and (S16), which does not depend ons, the S∗ ℓ’s are independent with marginal la...

    2020

  12. [440]

    James, L. F. (2017, October). Bayesian Poisson calculus for latent feature modeling via gener- alized Indian Buffet Process priors.The Annals of Statistics 45(5), 2016–2045. James, L. F., J. Lee, and A. Pandey (2023). Bayesian analysis of generalized hierarchical Indian buffet...

    2023 arXiv

  13. [449]

    De Iorio, and J

    Franzolini, B., M. De Iorio, and J. Eriksson (2023). Conditional partial exchangeability: a probabilistic framework for multi-view clustering.arXiv preprint arXiv:2307.01152. Ghahramani, Z. and T. Griffiths (2005). Infinite latent feature models and the indian buffet process. ...

    2023 arXiv

  14. [500]

    Favaro, and L

    Bacallado, S., S. Favaro, and L. Trippa (2013). Bayesian nonparametric analysis of reversible markov chains.The Annals of Statistics, 870–896. Baccelli, F., B. Błaszczyszyn, and M. Karray (2020). Random measures, point processes, and stochastic geometry.HAL preprint available ...

    2013

  15. [599]

    Ord, K. (1978). How many trees in a forest?The Mathematical Scientist 3, 23–33. Orlitsky, A., A. T. Suresh, and Y. Wu (2016). Optimal prediction of the number of unseen species. Proceedings of the National Academy of Sciences 113(47), 13283–13288. Petralia, F., V. Rao, and D. ...

    1978

  16. [836]

    Broderick, T., A. C. Wilson, and M. I. Jordan (2018). Posteriors, conjugacy, and exponential families for completely random measures.Bernoulli 24(4B), 3181 –

    2018

  17. [900]

    Prünster, I. (2002). Random probability measures derived from increasing additive processes and their application to Bayesian statistics. Ph. D. thesis, University of Pavia. Regazzini, E. (1978). Intorno ad alcune questioni relative alla definizione del premio secondo la teori...

    2002

  18. [1099]

    Bayesian calculus and predictive characterizations of extended feature allocation models

    Zabell, S. L. (1995). Characterizing markov exchangeable sequences. Journal of Theoretical Probability 8, 175–178. Zabell, S.L.(2005). Symmetry and its discontents. CambridgeStudiesinProbability, Induction, and Decision Theory. Cambridge University Press, New York. Essays on t...

    1995

  19. [1283]

    Rediscoveryofgood–turingestimatorsviabayesian nonparametrics

    Favaro, S., B.Nipoti, andY.W.Teh(2016). Rediscoveryofgood–turingestimatorsviabayesian nonparametrics. Biometrics 72(1), 136–145. Fong, E., C. Holmes, and S. G. Walker (2023). Martingale posterior distributions.J. R. Stat. Soc. Ser. B. Stat. Methodol. 85(5), 1357–1391. Fortini,...

    2016

  20. [1291]

    Prünster, and S

    Lijoi, A., I. Prünster, and S. G. Walker (2008). Investigating nonparametric priors with gibbs structure. Statistica Sinica, 1653–1668. Lo, A. Y. (1991). A characterization of the dirichlet process. Statistics & Probability Let- ters 12(3), 185–187. Lui, A., J. Lee, P. F. Thal...

    2008 arXiv

  21. [1448]

    Argiento, F

    32 Beraha, M., R. Argiento, F. Camerlenghi, and A. Guglielmi (2023). Bayesian mixtures models with repulsive and attractive atoms.arXiv preprint arXiv:2302.09034. Beraha, M. and J. E. Griffin (2023). Normalised latent measure factor models.Journal of the Royal Statistical Soci...

    2023 arXiv

  22. [1799]

    James, L. F. (2006). Poisson calculus for spatial neutral to the right processes.The Annals of Statistics 34(1), 416 –

    2006

  23. [2324]

    Hough, J. B., M. Krishnapur, Y. Peres, and B. Viràg (2006). Determinantal processes and independence. Probability Surveys 3, 206–229. Hu, Y., K. Zhai, S. Williamson, and J. Boyd-Graber (2012). Modeling images using transformed Indian buffet processes. InInternational Conferenc...

    2006 arXiv

  24. [3221]

    Sufficientness

    Camerlenghi, F. and S. Favaro (2021). On Johnson’s “Sufficientness” postulates for feature- sampling models. Mathematics 9(22). Camerlenghi, F., S. Favaro, L. Masoero, and T. Broderick (2023). Scaled process priors for Bayesian nonparametric estimation of the unseen genetic va...

    2021

Pith tools

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