Interval-censored Hawkes processes
Authors: Marian-Andrei Rizoiu, Alexander Soen, Shidi Li, Pio Calderon, Leanne J. Dong, Aditya Krishna Menon, Lexing Xie
JMLR 2022 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We verify our models through empirical testing on synthetic data and real-world data. We find that our MBPP outperforms HIP on real-world datasets for the task of popularity prediction. |
| Researcher Affiliation | Collaboration | Marian-Andrei Rizoiu EMAIL University of Technology Sydney Ultimo NSW 2007, Australia Alexander Soen EMAIL The Australian National University Canberra ACT 2601, Australia Shidi Li EMAIL The Australian National University Canberra ACT 2601, Australia Pio Calderon EMAIL University of Technology Sydney Ultimo NSW 2007, Australia Leanne J. Dong EMAIL Concordia University Montr eal, Canada Aditya Krishna Menon EMAIL Google Research Lexing Xie EMAIL The Australian National University Canberra ACT 2601, Australia |
| Pseudocode | No | The paper describes a simulation algorithm in Section 7.1, but it is presented as a descriptive list of steps rather than a formally structured pseudocode or algorithm block. |
| Open Source Code | No | The paper discusses various software tools that implement Hawkes processes (e.g., THAP, Po PPy, pyhawkes, tick, evently) in Section 9.3, but it does not provide an explicit statement or a link to the source code for the methodology developed in this specific paper. |
| Open Datasets | Yes | We use the ACTIVE dataset introduced by Rizoiu et al. (2017b) and further studied in (Wu et al., 2019, 2020). |
| Dataset Splits | Yes | For Scenarios A-C, we consider 1,000 event sequences, which are further split into 50 groups to estimate parameters. For Scenarios D-F, we consider 10,000 sequences, also split into 50 groups to be jointly fit. ... When we are testing intervalcensored settings, we use uniform observation intervals with 5, 10, 15, 30, 60, and 100 observation periods for example, with 10 observation intervals we calculate volumes over intervals (0, 3], (3, 6], . . . , (27, 30]. ... We fit model parameters on the first 90 days and we use the later 30 days for evaluation. |
| Hardware Specification | No | We thanks the Ne CTAR Research Cloud for providing computational resources, an Australian research platform supported by the National Collaborative Research Infrastructure Strategy. Furthermore, this research was undertaken with the assistance of resources and services from the National Computational Infrastructure (NCI), which is supported by the Australian Government. |
| Software Dependencies | No | The paper does not provide specific software dependencies with version numbers for its implementation. It mentions other tools in the related work section but not the exact software used by the authors. |
| Experiment Setup | Yes | The third augmentation deals with the non-convexity of the loss surface (see Appendix D for more details), which can cause the gradient descent minimizer to get stuck into local optima. We address this by refitting each video 10 times, starting with randomly selected initial parameters. We chose as the final parameters the combination with the lowest loss. ... augmented exogenous intensity for each day i is: s[i] = γ Jt = 0K + ν Jt > 0K + µ [#Tweets on day i], (54) where µ corresponds to the scaling constant of the observed exogenous influence; and γ and ν correspond to parameters accounting for unobserved external influence. |