yule-buaa / dnntsp

codes of DNNTSP model for Temporal Sets Prediction

Predicting Temporal Sets with Deep Neural Networks (DNNTSP)

DNNTSP is a general neural network architecture that could make prediction on temporal sets.

Please refer to our KDD 2020 paper “Predicting Temporal Sets with Deep Neural Networks” for more details.

Project Structure

The descriptions of principal files in this project are explained as follows:

• ./model/
  • weighted_graph_conv.py: codes for the Element Relationship Learning component (i.e. weighted GCN on dynamic graphs)
  • masked_self_attention.py and aggregate_nodes_temporal_feature.py: codes for the Attention-based Temporal Dependency Learning component (i.e. masked self-attention and weighted aggregation of temporal information)
  • global_gated_update.py: codes for the Gated Information Fusing component (i.e. gated updating mechanism)
• ./train/
  • train_model.py and train_main.py: codes for training models
• ./test/
  • testing_model.py: codes for evaluating models
• ./utils/: containing useful files that are required in the project (e.g. data loader, metrics calculation, loss function, configurations)
• ./data/: processed datasets are under in this folder. Original datasets could be downloaded as follows:
  • TaFeng
  • DC
  • TaoBao
  • TMS
• ./save_model_folder/ and ./runs/: folders to save models and outputs of tensorboardX respectively
• ./results/: folders to save the evaluation metrics for models.

Parameter Settings

Please refer to our paper for more details of parameter settings. Hyperparameters could be found in ./utils/config.json and you can adjust them when running the model.

How to use

• Training: after setting the parameters, run train_main.py file to train models.
• Testing: figure out the path of the specific saved model (i.e. variable model_path in ./test/testing_model.py) and then run testing_model.py file to evaluate models.

Principal environmental dependencies as follows:

• PyTorch 1.5.0
• dgl 0.5.2
• tqdm
• numpy
• pandas
• sklearn
• tensorboardX

Citation

Please consider citing the following paper when using our code.

@inproceedings{DBLP:conf/kdd/YuSDL0L20,
  author    = {Le Yu and
               Leilei Sun and
               Bowen Du and
               Chuanren Liu and
               Hui Xiong and
               Weifeng Lv},
  title     = {Predicting Temporal Sets with Deep Neural Networks},
  booktitle = {{KDD} '20: The 26th {ACM} {SIGKDD} Conference on Knowledge Discovery
               and Data Mining, Virtual Event, CA, USA, August 23-27, 2020},
  pages     = {1083--1091},
  year      = {2020}
}

dnntsp's Issues

PyTorch Geometric Temporal

HI there,

Would you love to contribute this to the library?

We would love to see this and of course it would be cited in the deisgn paper.

Bests,

Benedek

CUDA out of mem error on large temporal sets

I tried running the code with batch size 1 on a dataset where number of items are in order of 100-1000 across all sets in a datapoint. It fails with CUDA out of mem error. Specifically the code fails at:
https://github.com/yule-BUAA/DNNTSP/blob/master/model/weighted_graph_conv.py#L35

Is there any way to optimize the size of the edge_weights considering the fact that if we have say 600 nodes then size of edge weight comes around 30x360000 where 30 is the number of sets in a single datapoint.

Do you suggest optimization at a different step?