Stochastic Deep Gaussian Processes over Graphs

This is the official implementation and results of the DGPG model proposed in our paper: Stochastic Deep Gaussian Processes over Graphs, (NeurIPS 2020).

Prerequests

our implementation is mainly based on tensorflow 1.x and gpflow 1.x:

python 3.x (3.7 tested)
conda install tensorflow-gpu==1.15
pip install keras==2.3.1
pip install gpflow==1.5
pip install gpuinfo

Besides, some basic packages like numpy are also needed. It's maybe easy to wrap the codes for TF2.0 and GPflow2, but it's not tested yet.

Specification

Source code and experiment result are both provided. Unzip two archive files before using experiment notebooks.

Files

• dgp_graph/: cores codes of the DGPG model.
  • impl_parallel.py: a fast node-level computation parallelized implementation, invoked by all experiments.
  • my_op.py: some custom tensorflow operations used in the implementation.
  • impl.py: a basic loop-based implementation, easy to understand but not practical, leaving just for calibration.
• data/: datasets.
• doubly_stochastic_dgp/: codes from repository DGP
• compatible/: codes to make the DGP source codes compatible with gpflow1.5.
• gpflow_monitor/: monitoring tool for gpflow models, from this repo.
• demo_city45.ipynb: jupyter notebooks for city45 dataset experiment.
• experiments.zip: jupyter notebooks for other experiments.
• results.zip: contains original jupyter notebooks results. (exported as HTML files for archive)
• run_toy.sh: shell script to run additional experiment.
• toy_main.py: code for additional experiment (Traditional ML methods and DGPG with linear kernel).

Experiments

The experiments are based on python src files and demonstrated by jupyter notebooks. The source of an experiment is under directory src/experiments.zip and the corresponding result is exported as a static HTML file stored in the directory results.zip. They are organized by dataset names:

1. Synthetic Datasets
   • demo_toy_run1.ipynb
   • demo_toy_run2.ipynb
   • demo_toy_run3.ipynb
   • demo_toy_run4.ipynb
   • demo_toy_run5.ipynb
2. Small Datasets
   • demo_city45.ipynb
   • demo_city45_linear.ipynb (linear kernel)
   • demo_city45_baseline.ipynb (traditional regression methods)
   • demo_etex.ipynb
   • demo_etex_linear.ipynb
   • demo_etex_baseline.ipynb
   • demo_fmri.ipynb
   • demo_fmri_linear.ipynb
   • demo_fmri_baseline.ipynb
3. Large Datasets (traffic flow prediction)
   • LA
     • demo_la_15min.ipynb
     • demo_la_30min.ipynb
     • demo_la_60min.ipynb
   • BAY
     • demo_bay_15min.ipynb
     • demo_bay_30min.ipynb
     • demo_bay_60min.ipynb

Just run these notebooks to get the first-hand experience. Feel free to build your own experiments using core python codes.

Citations

If you find this repository, e.g., the code and the datasets, useful in your research, please cite the following paper:

@article{li2020stochastic,
  title={Stochastic Deep Gaussian Processes over Graphs},
  author={Li, Naiqi and Li, Wenjie and Sun, Jifeng and Gao, Yinghua and Jiang, Yong and Xia, Shu-Tao},
  journal={Advances in Neural Information Processing Systems},
  volume={33},
  year={2020}
}

Contanct us via github issue or email if you are interested in more details not covered here. Discussion and further extension is also welcoming