Fast and Flexible Temporal Point Processes with Triangular Maps

Authors: Oleksandr Shchur, Nicholas Gao, Marin Biloš, Stephan Günnemann

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 6 Experiments", "Table 1: Average test set NLL on synthetic and real-world datasets (lower is better).", "Table 2: MMD between the hold-out test set and the generated samples (lower is better).
Researcher Affiliation Academia Oleksandr Shchur, Nicholas Gao, Marin Biloš, Stephan Günnemann Technical University of Munich, Germany {shchur,gaoni,bilos,guennemann}@in.tum.de
Pseudocode No The paper describes algorithms and methods but does not include any clearly labeled pseudocode or algorithm blocks.
Open Source Code Yes Code and datasets are available under www.daml.in.tum.de/triangular-tpp
Open Datasets Yes We use 6 synthetic datasets from Omi et al. [10]: Hawkes1&2 [7], self-correcting (SC) [16], inhomogeneous Poisson (IPP), renewal (RP) and modulated renewal (MRP) processes. ... We also consider 7 real-world datasets: PUBG (online gaming), Reddit-Comments, Reddit-Submissions (online discussions), Taxi (customer pickups), Twitter (tweets) and Yelp1&2 (check-in times). See Appendix D for more details.
Dataset Splits Yes We partitioned the sequences in each dataset into train/validation/test sequences (60%/20%/20%). We trained the models by minimizing the NLL of the train set using Adam [57]. We tuned the following hyperparameters: ... We used the validaiton set for hyperparameter tuning, early stopping and model development.
Hardware Specification Yes We used a machine with an Intel Xeon E5-2630 v4 @ 2.20 GHz CPU, 256GB RAM and an Nvidia GTX1080Ti GPU.
Software Dependencies No The paper mentions 'Py Torch [53]' for implementation but does not specify its version number, nor does it list versions for other software dependencies.
Experiment Setup Yes We tuned the following hyperparameters: L2 regularization {0, 10 5, 10 4, 10 3}, number of spline knots {10, 20, 50}, learning rate {10 3, 10 2}, hidden size {32, 64} for RNN, number of blocks {2, 4} and block size {8, 16} for Tri TPP.