Node-level Graph Regression with Deep Gaussian Process Models
our implementation is mainly based on
tensorflow 1.x and
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.
Source code and experiment result are both provided. Unzip two archive files before using experiment notebooks.
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.
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.
GRN inference: code and data for the GRN inference experiment.
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).
ER-0.1.ipynb: example script for analyzing time-varying graph structures.
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:
- Synthetic Datasets
For theoretical analysis.
For graph signal analysis on time-varying graphs.
- Small Datasets
demo_city45_baseline.ipynb(traditional regression methods)
- Large Datasets (traffic flow prediction)