Share your thoughts, 1 month free Claude Pro on usSee more
WorkDL logo mark

Fast and Flexible Temporal Point Processes with Triangular Maps

About

Temporal point process (TPP) models combined with recurrent neural networks provide a powerful framework for modeling continuous-time event data. While such models are flexible, they are inherently sequential and therefore cannot benefit from the parallelism of modern hardware. By exploiting the recent developments in the field of normalizing flows, we design TriTPP -- a new class of non-recurrent TPP models, where both sampling and likelihood computation can be done in parallel. TriTPP matches the flexibility of RNN-based methods but permits orders of magnitude faster sampling. This enables us to use the new model for variational inference in continuous-time discrete-state systems. We demonstrate the advantages of the proposed framework on synthetic and real-world datasets.

Oleksandr Shchur, Nicholas Gao, Marin Bilo\v{s}, Stephan G\"unnemann• 2020

Related benchmarks

TaskDatasetResultRank
Temporal Point Process modelingHawkes2 synthetic (test)
NLL-1.939
25
Temporal Point Process modelingHawkes1 synthetic (test)
Negative Log-Likelihood-1.51
25
Univariate Temporal Point Process modelingRenewal1 synthetic (test)
NLL-1.821
14
Univariate Temporal Point Process modelingRenewal2 synthetic (test)
NLL-1.91
14
Univariate Temporal Point Process modelingSelf-correcting synthetic (test)
Negative Log-Likelihood (NLL)-1.328
14
Event Sequence ForecastingYelp1 (test)
Wasserstein Distance0.03
13
Event Sequence ForecastingTwitter (test)
Wasserstein Distance0.01
13
Event Sequence ForecastingPUBG (test)
Wasserstein Distance0.03
13
Event Sequence ForecastingReddit-S (test)
Wasserstein Distance0.09
13
Event Sequence ForecastingYelp-2 (test)
Wasserstein Distance0.04
13
Showing 10 of 35 rows

Other info

Follow for update