Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
Scalable and Adaptive Variational Bayes Methods for Hawkes Processes
Authors: Deborah Sulem, Vincent Rivoirard, Judith Rousseau
JMLR 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Finally, through an extensive set of numerical simulations, we demonstrate that our method is able to adapt to the dimensionality of the parameter of the Hawkes process, and is partially robust to certain types of model misspecification. In addition to being theoretically valid in the asymptotic regime, we show that our algorithm performs very well in practice. In an extensive set of simulations, we observe that it is scalable to large data sets. Finally, we report in Section 6 the results of an in-depth simulation study. |
| Researcher Affiliation | Academia | D eborah Sulem EMAIL Barcelona School of Economics Universitat Pompeu Fabra Vincent Rivoirard EMAIL Ceremade, CMRS, UMR 7534 Universit e Paris-Dauphine, PSL University Judith Rousseau EMAIL Department of Statistics University of Oxford |
| Pseudocode | Yes | Algorithm 1: Mean-field variational inference algorithm in a fixed model Algorithm 2: Fully-adaptive mean-field variational inference Algorithm 3: Two-step adaptive mean-field variational inference Algorithm 4: Gibbs sampler in the sigmoid Hawkes model with data augmentation |
| Open Source Code | No | The paper does not contain any explicit statement about the release of source code for the described methodology, nor does it provide any links to a code repository. |
| Open Datasets | No | In each setting, unless specified otherwise, we sample one observation of a Hawkes process with dimension K, link functions (φk)k and parameter f0 = (ν0, h0) on [0, T], using the thinning algorithm of Adams et al. (2009). The paper uses simulated data for its experiments and does not provide access information for any publicly available or open dataset. |
| Dataset Splits | No | The paper uses simulated data for its evaluation. There is no mention of splitting a dataset into training, validation, or test sets because the experiments are based on generating new data, not on evaluating pre-existing datasets that would typically require such splits. |
| Hardware Specification | Yes | In an extensive set of simulations, we observe that it is scalable to large data sets. In particular, on a desktop with only 8 cores, our algorithm takes less than 8 hours to run for processes with 64 dimensions and more than 88000 events. |
| Software Dependencies | Yes | To compute the (true) posterior distribution, we run a Metropolis-Hasting (MH) sampler implemented via the Python package Py MC41 |
| Experiment Setup | Yes | In all simulations, we set the memory parameter as A = 0.1. In each of the nine settings, we set T = 500 and in Table 1, we report the corresponding number of events and excursions observed in each scenario and model. We consider a normal prior on HD0 histo such that w11 N(0, σ2I), and for ν1, ν1 N(0, σ2), with σ = 5.0. To compute the (true) posterior distribution, we run a Metropolis-Hasting (MH) sampler implemented via the Python package Py MC41 with 4 chains, 40 000 iterations, and a burn-in time of 4000 iterations. We run 4 chains for 40 000 iterations for the MH sampler, 3000 iterations of the Gibbs sampler, and use our early-stopping procedure for the mean-field variational algorithm. Here, we fix the dimensionality of h11 to J = 2D0 = 4. In our adaptive variational algorithm, we set a maximum histogram depth D1 = 5 for K = 1, and D1 = 4 for K = 2. |