Is Score Matching Suitable for Estimating Point Processes?

Score matching estimators have gained widespread attention in recent years partly because they are free from calculating the integral of normalizing constant, thereby addressing the computational challenges in maximum likelihood estimation (MLE). Some existing works have proposed score matching estimators for point processes. However, this work demonstrates that the incompleteness of the estimators proposed in those works renders them applicable only to specific problems, and they fail for more general point processes. To address this issue, this work introduces the weighted score matching estimator to point processes. Theoretically, we prove the consistency of our estimator and establish its rate of convergence. Experimental results indicate that our estimator accurately estimates model parameters on synthetic data and yields results consistent with MLE on real data. In contrast, existing score matching estimators fail to perform effectively. Codes are publicly available at \url{https://github.com/KenCao2007/WSM_TPP}.

Paper

References (28)

Scroll for more · 16 remaining

Similar papers

Peer review

Reviewer X8YN7/10 · confidence 4/52024-07-04

Summary

EDIT: I have changed my score from a "weak accept" to an "accept." The paper studies the potential use of score matching (SM) to do inference with the model of interest is a Poisson point process or Hawkes processes. Because using SM in practice requires modifying the objective with integration by parts, the paper shows for many common Poisson and Hawkes processes a core assumption to make this modification possbile does not hold. In order to remedy the situation, the paper proposes a random weighted version of SM called WSM (along with an analog for Autoregressive Score Matching (ASM) called WASM), which reestablishes the validity of integration by parts for the aforementioned processes. The paper argues theoretically that their new estimators are consistent, and it provides empirical analyses to demonstrate that the new estimator outperforms the non-weighted variant and matches the performance of alternatives. It also makes an argument to select a weighting function based on some set of assumptions.

Strengths

The paper points out a fundamental issue with trying to use SM for general Poisson point processes and Hawkes processes, and in addition, it recommends a weighting scheme that patches up this problem based on some reasonable assumptions. This is a valuable observation, and the proposed remedy looks reasonable. There is ample empirical work to suggest that their tweak on SM and ASM does agree with the maximum likelihood estimator (MLE). The theory provided in the paper also argues that their weighted estimator is consistent if the model is well-specified. Originality: [+] + The authors have applied a weighting scheme to SM and ASM that permits them to use integration by parts for the standard objective and prove the consistency of their new estimator under suitable regularity conditions. Quality: [+] + The empirical work is thorough and does a good job checking the accuracy of their new estimator. Clarity: [+] + The presentation of the theory and empirical work is very neat and polished. + Assumptions are clearly laid out to ensure consistency of the new estimator. Significance: [+] + The new estimator provides the ability to do consistent semiparametric estimation for Poisson point and Hawkes processes with SM (instead of MLE), which provides a new tool for very large dimensional problems.

Weaknesses

There are a couple critiques of the paper: - While the idea of using a weighting function to enable integration by parts is a good one, it has been seen in other fields before (e.g. weighting Stein kernels in compact settings), and hence is not completely novel. - The argument in Section 5 outlining a weighting function is a great addition to the paper, but given that it is only optimizing $C_h$ and not $\Gamma(h, A, B)$, it is hard to know how close it is to the optimum. Originality: [-] - The only novel insight of the paper is that the trick of integration by parts, so often using in score matching, is not possible for general point processes, but it can be recovered by weighting the Fisher divergence a bit differently. It is not clear if this idea alone is sufficient for publication. Quality: [-] - The argument to choose their weighting function $h$ is definitely better than simply picking one. That being said, it still not clear how close to optimal their choice of $h$ is given they are only optimizing one term in the objective ($C_h$). The paper would be a bit stronger if they could include a different (natural) weighting function to show their new choice is an improvement. Clarity: [-] - Equation 1: Can we at least have a sentence explaining the paper will be slightly abusing notation by referring to $p(T)$ as the conditional distribution of $p$ given $N_t = |T|$?

Questions

[Q1] Given the choice of $h$ optimizes $C_h$ but ignores $\Gamma(h, A, B)$, how sensitive is $\Gamma$ to the choice of $h$? This could be demonstrated empirically or theoretically, but it would be good to have some analysis of this to argue why this weighting function is decent.

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

NA

Reviewer B3db5/10 · confidence 4/52024-07-13

Summary

This paper studies the use of score matching for estimation in point process models, and proves the incompleteness in the original score matching estimators due to the bounded support. Weighted score matching is use to address the issue and theoretical results are establish ed for the consistency and optimality of the proposed estimator. Experiments on different data demonstrate the effectiveness of the proposed estimator which yields results comparable to maximum likelihood estimation.

Strengths

1. This paper theoretically shows the limitation in the use of original score matching estimators for point processes, pointing out their incompleteness and provide a solution. 2. Theoretical analysis is given to support the claim and the convergence of the proposed method. 3. The paper is well-written and easy to follow.

Weaknesses

1. As acknowledged in the limitation section in the paper, existing approaches that adopt denoising score matching on the point process is not considered. Even with theoretical gaurantee, experimentally, such baselines would be necessary to include in order to fully demonstrate the superiority of the method. 2. For Hawkes processes, the problem cause by the bounded support can be solved by applying log-normalization to transform the bounded temporal domain into an unbounded one [1]. One could achieve this by removing the right bound and transform the $(0,+\infty)$ to $(-\infty,+\infty)$. Although the real data is observed within finite window, the effect of removing the right bound should be negligible as it only affect the last event. 3. Including more metrics on event prediction, such as MSE of event time, would be helpful, as the practical use of point process model is for prediction. [1] Lin, H., Wu, L., Zhao, G., Liu, P., & Li, S. (2022). Exploring generative neural temporal point process. Transactions on Machine Learning Research.

Questions

* Did the author observe any instability during the score matching training?

Rating

5

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

Included in weakness

Authorsrebuttal2024-08-12

We greatly thank the reviewer for your constructive feedback. > For the log-normalization, could you kindly provide the detailed derivation for obtaining $\hat \lambda_{\text{ASM,Log}}$ and $\hat \lambda _{\text{ASM,Log Interval}}$? Sure, we are more than delighted to discuss this with you. For log-normalization on timestamps, since the transformed variable $x_n$ has conditional pdf, $ p(x_n|\mathcal F_{x_{n-1}})=\lambda \exp\left[-\lambda \exp(x_n)\right]\exp(x_n). $ The conditional score and its partial derivative w.r.t. $x_n$ will be, $ \psi(x_n|\mathcal F_{x_{n-1}}) = -\lambda \exp(x_n) + 1, $ $ \frac{\partial}{\partial x_n}\psi_{\theta}(x_n|\mathcal F_{x_{n-1}})=-\lambda \exp(x_n). $ Now we perform ASM on the transformed variable $x_n$. Denote $\mathcal X = (x_1,\ldots, x_{N_T})^T$ as the vector of transformed timestamps, the implicit ASM objective (Eq. 13) for $\mathcal X$ will be, $ \mathcal J_\text{ASM}(\theta)=\mathbb E_{p(\mathcal X)}[\sum_{n=1}^{N_T}\frac{1}{2}\psi^2_\theta(x_n|\mathcal F_{x_{n-1}})+\frac{\partial}{\partial x_n}\psi_\theta(x_n|\mathcal F_{x_{n-1}})]. $ We plug in the conditional score we just computed into the above equation and get, $ \mathcal J_{\text{ASM}}(\theta) =\mathbb E_{p(\mathcal X)}[\sum_{n=1}^{N_T}\frac{1}{2}\exp(2x_n)\lambda^2-\sum_{n=1}^{N_T}2\exp(x_n)\lambda ]+C, $ where $C$ is a constant not containing $\theta$. The empirical objective based on samples $t_1^{(m)},\ldots t_{N_m}^{(m)}, m=1,\ldots M$ will be, $\hat J_\text{ASM}(\theta)=\frac{1}{M}\sum_{m=1}^M\sum_{n=1}^{N_m}[\frac{1}{2}\exp(2x_n^{(m)})\lambda^2-2\exp(x_n^{(m)})\lambda]. $ This is quadratic with respect to $\lambda$, and its minimal $\hat \lambda_{\text{ASM,Log}}$ can be computed analytically as $ \hat \lambda_{\text{ASM,Log}} = 2\frac{\sum_{m=1}^M\sum_{n=1}^{N_m}\exp(x_n^{(m)})}{\sum_{m=1}^M\sum_{n=1}^{N_m}\exp(2x_n^{(m)})}. $ For log transformation on time interval $y_n$, since the conditional pdf of $\tau_n = t_n-t_{n-1}$ is, $ p(\tau_n|\mathcal F_{\tau_{n-1}})=\lambda \exp(-\lambda \tau_n), $ thus the conditional pdf of $y_n = \log \tau_n$ will be, $ p(y_n|\mathcal F_{y_{n-1}})=\lambda \exp\left[-\lambda \exp(y_n)\right]\exp(y_n), $ it has the same analytical expression as that of $x_n$, except that $y_n$ and $x_n$ have different support. Therefore, following exactly the same steps as those of $x_n$, we can get the analytical expression of $\hat \lambda_{\text{ASM,Log, Interval}}$ in our rebuttal (with minor modification for adding superscript $(m)$). > Additionally, is it possible to express $\hat \lambda_{\text{AWSM,Log}}$ and $\hat \lambda_{\text{AWSM, Log Interval}}$ in closed form? Yes, it is possible. We take $\hat \lambda_{\text{AWSM, Log}}$ as an example. The derivation of $\hat \lambda_{\text{AWSM, Log Interval}}$ is basically the same. First, the implicit AWSM objective as Eq. 17 in our paper for $\mathcal X$ will be, $ \mathcal J_\text{AWSM}(\theta)=\mathbb E_{p(\mathcal X)}[\sum_{n=1}^{N_T}\frac{1}{2}\psi^2_\theta(x_n|\mathcal F_{x_{n-1}})h_n(\mathcal X)+\frac{\partial}{\partial x_n}\psi_\theta(x_n|\mathcal F_{x_{n-1}})h_n(\mathcal X)+\psi_\theta(x_n|\mathcal F_{x_{n-1}})\frac{\partial}{\partial x_n}h_n(\mathcal X)]. $ Plug in $\psi_\theta(x_n|\mathcal F_{x_{n-1}})$ and $\frac{\partial}{\partial x_n}\psi_\theta(x_n|\mathcal F_{x_{n-1}})$ into the empirical version of the above equation and get, $ \hat J_\text{AWSM}(\theta)=\frac{1}{M}\sum_{m=1}^M\sum_{n=1}^{N_m}[\frac{1}{2}\exp(2x_n^{(m)})h_n(\mathcal X^{(m)})\lambda^2-2\exp(x_n^{(m)})h_n(\mathcal X^{(m)})\lambda - \exp(x_n^{(m)})\frac{\partial}{\partial x_n}h_n(\mathcal X^{(m)})\lambda]. $ This is still quadratic w.r.t $\lambda$, and its minimal will be, $ \hat \lambda_{\text{AWSM, Log}}=\frac{\sum_{m=1}^M\sum_{n=1}^{N_m}[2\exp(x_n^{(m)})h_n(\mathcal X^{(m)})+\exp(x_n^{(m)})\frac{\partial}{\partial x_n}h_n(\mathcal X^{(m)})]}{\sum_{m=1}^M\sum_{n=1}^{N_m}\exp(2x_n^{(m)})h_n(\mathcal X^{(m)})}. $ Since we choose weight $h_n(\mathcal X)=(x_n-x_{n-1})(\log T-x_n)$ with derivative $\frac{\partial}{\partial x_n}h_n(\mathcal X)=\log T - 2x_n+x_{n-1}$. Plug them in the above equation, then we get a closed form of $\hat \lambda_{\text{AWSM, Log}} $.

Reviewer B3db2024-08-12

Thank you for the further derivation and clarification! For the log interval case, each $\tau_i$ should follow the same exponential distribution, so $\hat{\lambda}_{\text{ASM,Log Interval}}$ should approximate the correct $\lambda$ with a large sample size. Could you please clarify why this is considered incorrect? I might have misunderstood something.

Authorsrebuttal2024-08-12

We thank greatly for the reviewer for further engaging in the discussion. > For the log interval case, each $\tau_i$ should follow the same exponential distribution, so $\hat{\lambda}_{\text{ASM,Log Interval}}$ should approximate the correct $\lambda$ with a large sample size. This is not true. It is widely known that for an exponential distribution on $[0,\infty)$, it indeed has expectation $\frac{1}{\lambda}$ and second moments $\frac{2}{\lambda^2}$. Then the average of exponential distribution should be equal to $\frac{1}{\lambda}$ when sample is large. However, this is not the case when considering the time interval in a Poisson process on $[0,T]$. Considering the numerator of $\hat \lambda_{\text{AWSM,Log Interval}}$, we **cannot** draw the conclusion that, $ \lim_{M\rightarrow \infty}\mathbb E_{p}[\frac{\sum_{m=1}^M\sum_{n=1}^{N_m}\tau_n^{(m)}}{\sum_{m=1}^MN_m}]=\frac{1}{\lambda} $ We cannot use CLT or other things here since $\sum_{m=1}^M N_m$ is random and each $\tau_n^{(m)}$ does not necessarily have the same marginal distribution since their support is different. To illustrate this, we compute $\frac{\sum_{m=1}^M\sum_{n=1}^{N_m}\tau_n^{(m)}}{\sum_{m=1}^MN_m}$ for $M=10^{6}, T = 1$ under different $\lambda^{*}$ and show that the expectation of it is not equal to $\frac{1}{\lambda}$. | | $\lambda=1$ | $\lambda = 2$ | $\lambda = 4$ | $\lambda = 8$ | | ------------------------------------------------------------ | ----------- | ------------- | ------------- | ------------- | | $\frac{\sum_{m=1}^M\sum_{n=1}^{N_m}\tau_n^{(m)}}{\sum_{m=1}^MN_m}$ | 0.3693 | 0.2836 | 0.1886 | 0.1094 | Also for the denominator, its expectation is not equal or asymptotically equal to $\frac{2}{\lambda^2}$ when taking average. Therefore, $\hat{\lambda}_{\text{ASM,Log Interval}}$ does not approximate the correct $\lambda$ with a large sample size. We give a brief explanation as follows. For a stochastic process in a finite window $[0,T]$. Each $\tau_i$ has support $[0, T-\sum_{i\leq n}\tau_i]$ instead of $[0,\infty)$. Therefore, firstly, each $\tau_i$ has different support and does not follow the same distribution. Secondly, for each $\tau_i$ , it has support different from that of the exponential distribution, which is $[0, \infty)$. In conclusion, when discussing a point process on a finite time window, the right bound's effect must be considered. And the case is different from sampling a fixed length time sequence over an infinite time window.

Reviewer Se8E6/10 · confidence 4/52024-07-15

Summary

The paper considers the problem of utilizing score-matching approaches for point processes. The main motivations for the paper are the Poisson and Hawkes processes. Both of these model the repeated occurrences of events over a finite time interval $[0, T]$. The Poisson process models the occurrence of an event within a fixed infinitesimal interval with a fixed homogenous rate $\lambda$ while a Hawkes process is a self-excitation process where past events increase the likelihood of future ones. Score-matching is a natural approach in settings where the computation of the normalizing constant of a parameterized model are challenging to compute. However, to ensure traceability, one utilizes an implicit approach where the score matching objective is approximated by a tractable alternative which may be easily computed through the derivatives of the approximating model. The main contribution of this work is the identification of simple scenarios, even for the simple setting of the generalized Poisson process, where the score function optimization function of prior approaches fails as the regularity condition underlying them does not apply. The paper derives an expression showcasing the bias (Proposition 3.1) and is related to the evaluation of the estimated score function (for fixed number of occurrences) at the end points of the interval of interest ($t = 0$ and $t = T$). The paper then proposes an alternative where a weight function $h$ is used to augment the score function loss which ensures that these biasing terms evaluate to $0$. Furthermore, these functions may be chosen independently of the parameters of the underlying model. In their experiments, it is shown that the use of such functions still satisfies desirable properties such as the optima corresponding to the true values of the parameters while also allowing for a tractable score function formulation which avoids the bias incurred by prior approaches. This is a natural approach which suggests an avenue towards improving the performance of such models. Overall, the problem considered in the paper and the solution they identify are interesting and likely relevant to the NeurIPS community. However, I do have concerns regarding the writing of the paper. For example, there are several assumptions on the weight function in Equation 9 but it is not clear which properties are used in the proof of Proposition 3.3. For instance, Line 400 features an explicit expression for $h_1$ and $h_N$ while the expression in Equation 9 features none of these. It is my understanding that the assumptions in Equation 9 ensure that the biasing terms equate to $0$ for these elements. Furthermore, I believe that under the assumptions in Equation 9, both terms appearing in the display between Lines 399 and 400 are $0$. Please clarify whether this is in fact the case. The same issues persist for the proofs of the results in Section 4. I would like clarification from the authors on these points before updating my review. ************************************************************************************************************************************************ I thank the authors for their response and have amended my review accordingly.

Strengths

See main review

Weaknesses

See main review

Questions

See main review

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

See main review

Authorsrebuttal2024-08-09

Thanks

Thank you very much for your thoughtful review and for taking the time to carefully consider our responses. We greatly appreciate your kind words and are pleased that you find our work to be well-presented, thorough, and novel. We will certaintly correct the proofs in the updated manuscript, as you suggested. Your feedback is valuable in improving the quality of our work, and we are grateful for your support. Thank you once again for your constructive feedback and for increasing your rating.

Area Chair UHVE2024-08-07

Dear Reviewers, Now that the rebuttal period has ended, please take a moment to review the authors' responses to your initial reviews. Your engagement is crucial for: Ensuring a fair evaluation process Improving the quality of final reviews Fostering meaningful scientific discourse If you have any questions or need clarification, please post follow-up comments. Your continued dedication to this process is greatly appreciated. Best regards, AC

Reviewer X8YN2024-08-09

Thanks for your reply. I appreciate that you have done some extra work to demonstrate the choice of $h^0$ is better than other obvious alternatives. I am inclined to accept this paper.

Authorsrebuttal2024-08-09

Thanks

Thank you very much for your recognition and support. We will include the additional comparison of different weight functions in the camera-ready version. Thank you once again for your constructive feedback and for increasing your rating to accept.

Reviewer B3db2024-08-12

Thank you sincerely for the additional experiments on denoising score matching and clarifications! For the log-normalization, could you kindly provide the detailed derivation for obtaining $\hat{\lambda}\_{\text{ASM,Log}}$ and $\hat{\lambda}\_{\text{ASM,Log Interval}}$? Additionally, is it possible to express $\hat{\lambda}\_{\text{AWSM,Log}}$ and $\hat{\lambda}\_{\text{AWSM,Log Interval}}$ in closed form?

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC