This work introduces a Graph Convolutional Hawkes Process (GCHP) method incorperating marked point processes model to deal with predictions on attributed event sequences.

Firstly, our model transforms the marked event sequences into temporal similarity graph with feature matrix. Then, we feed the data into the GCHP model.

The following python library dependencies need to be installed:

• pandas
• numpy
• torch (with CUDA) >= 1.7 (detailed installation documents can be found in https://pytorch.org/)

• ATM
• Weeplace
• IPTV
• HawkesProcess_synthetic (synthetic dataset)

To quickly run the model on ATM dataset, just run:

python train.py --dataset ATM

You may run GCHP on other datasets with different hyperparameter settings: epochs, batch_size, memory_size, theta, eta, hidden, etc. Here we show an example of training on IPTV:

python train.py --dataset IPTV --batch_size 50000 --memory_size 10 --epochs 200 --theta 1.0 --eta 1.0 --hidden 50 --record True --loss_type likelihood

--loss_type likelihood

denotes the likelihood ratio loss proposed in our paper.

@incollection{tianbo2021gchp,
 author = {Tianbo Li, Tianze Luo, Yiping Ke, Sinno Jialin Pan},
 booktitle = {Proceedings of the 27th ACM SIGKDD International Conference on Knowledge Discovery \& Data Mining},
 pages = {},
 title = {Mitigating Performance Saturation in Neural Marked PointProcesses: Architectures and Loss Functions},
 year = {2021}
}