AAAI 2023 Hierarchical Contrastive Learning for Temporal Point Processes
In this work, we develop a novel hierarchical contrastive (HCL) learning method for temporal point processes, which provides a new regularizer of maximum likelihood estimation. In principle, our HCL considers the noise contrastive estimation (NCE) problem at the event-level and that at the sequence-level jointly. Given a sequence, the event-level NCE maximizes the probability of each observed event given its history while penalizing the conditional probabilities of the unobserved events. At the same time, we generate positive and negative event sequences from the observed sequence and maximize the discrepancy between their likelihoods through the sequence-level NCE. Instead of using time-consuming simulation methods, we generate the positive and negative sequences via a simple but efficient model-guided thinning process.
If you use this code as part of any published research, please acknowledge the following paper:
@article{wang2023hierarchical,
title={Hierarchical Contrastive Learning for Temporal Point Processes},
author={Wang, Qingmei and Cheng, Minjie and Yuan, Shen and Xu, Hongteng},
year={2023}
}
Here are the instructions to use the code base
This code is written in python. To use it you will need:
- PyTorch == 1.10.0
- Python == 3.9.0
Downloadable from this Google Drive link
python main.py -batch_size 4 -epoch 50 -model 'MLE' -save_label 'MLE + Reg' -data_folder 'tpp-data/data_hawkes' -w_mle 1 -w_dis 1 -w_cl1 0 -w_cl2 0 -seed 12
python main.py -batch_size 4 -superpose -epoch 50 -model 'MLE' -save_label 'MLE + DA' -data_folder 'tpp-data/data_hawkes' -w_mle 1 -w_dis 1 -w_cl1 0 -w_cl2 0 -seed 12
python main.py -batch_size 4 -model 'MLE' -save_label 'Dis' -data_folder 'tpp-data/data_hawkes' -epoch 50 -w_mle 0 -w_dis 1 -w_cl1 0 -w_cl2 0 -seed 12
python main.py -batch_size 4 -num_neg 20 -ratio_remove 0.4 -model 'HCL' -save_label 'HCL+MLE ' -data_folder 'tpp-data/data_hawkes' -epoch 50 -w_mle 1 -w_dis 1 -w_cl1 1 -w_cl2 1 -seed 12
python main.py -batch_size 4 -num_neg 20 -ratio_remove 0.4 -model 'HCL' -save_label 'HCL+MLE ' -data_folder 'tpp-data/data_hawkes' -epoch 50 -w_mle 1 -w_dis 1 -w_cl1 1 -w_cl2 0 -seed 12
python main.py -batch_size 4 -num_neg 20 -ratio_remove 0.4 -model 'HCL' -save_label 'HCL+MLE ' -data_folder 'tpp-data/data_hawkes' -epoch 50 -w_mle 1 -w_dis 1 -w_cl1 0 -w_cl2 1 -seed 12
tpp-data is the dataset.
Learning is the learning methods chosen for the training, including mle, hcl.
TPPSis the model chosen for the backbone of training.
num_neg is the number of negative sequence for contrastive learning. The default value of Hawkes dataset is 20.
wcl1 corresponds to the weight of event level contrastive learning loss. The default value is 1.
wcl2 corresponds to the weight of event level contrastive learning loss. The default value is 1.
ratio_remove corresponds to the ratio of removing events of per sequence when generate negative and positive sequence . The default value is 0.4.