Exploring Generative Neural Temporal Point Process

Authors: Haitao Lin, Lirong Wu, Guojiang Zhao, Liu Pai, Stan Z. Li

TMLR 2022 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments have been conducted to illustrate the improved predictive capability of GNTPP with a line of generative probabilistic decoders, and performance gain from the revised attention.
Researcher Affiliation Academia Haitao Lin EMAIL CAIRI, Westlake University Zhejiang University Lirong Wu EMAIL CAIRI, Westlake University Zhejiang University Guojiang Zhao EMAIL CAIRI, Westlake University Carnegie Mellon University Pai Liu EMAIL School of Engineering, Westlake University Stan Z. Li EMAIL CAIRI, Westlake University
Pseudocode Yes Algorithm 1 Training for each timestamp ti > ti 1 in temporal point process in TCDDM 1: Input: Observation time interval τi and historical encoding hi 1 2: repeat 3: Initialize k Uniform(1, . . . , K) and ϵ N(0, 1) 4: Take gradient step on θ ϵ ϵθ( αkτi + 1 αkϵ, hi 1, k) 2 5: until converged Algorithm 2 Sampling ˆti > ti 1 via Langevin dynamics Input: noise ˆτ K i N(0, 1) and historical encoding hi 1 for k = K to 1 do if k > 1 then z N(0, 1) else z = 0 end if ˆτ k 1 i = 1 αk (ˆτ k i βk 1 αk ϵθ(ˆτ k i , hi 1, k)) + Σθz end for Return: ˆti = ˆτ 0 i + ti 1
Open Source Code Yes Our codebase including all the methods given in Section. 5.1.1 is open in https://github.com/BIRD-TAO/GNTPP.
Open Datasets Yes We use a complex synthetic dataset which is simulated by Hawkes process of five types of events with different impact functions (Appendix B.1.) and 4 real-world datasets containing event data from various domains: MOOC (user interaction with online course system), Retweet (posts in social media), Stack Overflow (question-answering badges in website), Yelp (check-ins to restaurants).
Dataset Splits Yes The dataset is split into 20% ratio for testing, 80% ratio for training with 20% in training set as validation set used for parameter tuning.
Hardware Specification Yes The test is implemented on a single Nvidia-V100(32510MB).
Software Dependencies No The paper mentions generating synthetic datasets using the 'tick' package (Bacry et al., 2017), but does not specify a version number for this or any other software dependency.
Experiment Setup Yes In the training process, hyper-parameters of every model are tuned in the range of learning rate : {1 10 3, 5 10 4, 1 10 4}, embedding size : {8, 16, 32}, layer number : {1, 2, 3}, where embedding size is the dimension of historical encoding, i.e. D. The hyper-parameters are tuned on validation set. The maximum training epoch number is set as 100, and early stopping technique is used based on values of loss on validation set.