Neural Spectral Marked Point Processes

Authors: Shixiang Zhu, Haoyun Wang, Zheng Dong, Xiuyuan Cheng, Yao Xie

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental This section presents experimental results on both synthetic and real data and compares them with several state-of-the-arts, including (i) standard Hawkes process with an exponentially decaying kernel function (Hawkes) (Hawkes, 1971); (ii) recurrent marked temporal point processes (RMTPP) (Du et al., 2016); and (iii) Neural Hawkes process (NH) (Mei & Eisner, 2017).
Researcher Affiliation Academia Shixiang Zhu , Haoyun Wang , Zheng Dong , Xiuyuan Cheng , Yao Xie Georgia Institute of Technology Duke University
Pseudocode Yes Algorithm 1: Stochastic gradient-based learning algorithm; Algorithm 2: Monte Carlo estimation for the integral of conditional intensity in (6); Algorithm 3: Efficient thinning algorithm for simulating point process
Open Source Code Yes Codes to reproduce the experimental results are publicly available1. 1https://github.com/meowoodie/Neural-Spectral-Marked-Point-Processes
Open Datasets Yes Northern California Earthquake Data Center. UC Berkeley Seismological Laboratory. Dataset (2014) that comes from broadband, short period, strong motion seismic sensors, GPS, and other geophysical sensors.
Dataset Splits No The paper specifies training and testing splits but does not explicitly mention a separate validation set split. For synthetic data: 'We fit the models using 80% of the synthetic data set and the remaining 20% as the testing set.' For real data: 'We fit the models using 80% of the data set and the remaining 20% as the testing set.'
Hardware Specification Yes All experiments are performed on Google Colaboratory (Pro version) with 12GB RAM and dual-core Intel processors, which speed up to 2.3 GHz (without GPU).
Software Dependencies No The paper mentions using the 'Adam optimizer Kingma & Ba (2014)' and 'Soft Plus' activation function, but it does not specify version numbers for general software dependencies like Python, PyTorch, TensorFlow, or specific libraries beyond their names.
Experiment Setup Yes We set the rank R = 5 in our setting. We consider the shared network that summarizes input data into a hidden embedding to be a fully connected three-layer network and the sub-network to be a fully connected network with two hidden layers. The width of the hidden layers in the shared network is n = 128, and the width of the input layers in sub-networks (or the output layer in the shared network) is p = 10. We adopt the Soft Plus f(x) = 1/ log(1 + exp(x)) as the activation function of each layer in the network. To learn the model s parameters, we adopt the Adam optimizer Kingma & Ba (2014) with a constant learning rate of 10 2 and the batch size is 32.