This repository contains codes that produced the results in the paper Gaussian Processes on Graphs via Spectral Kernel Learning
We recommend starting with the gma
folder which contains codes for learning on synthetic data.
main.py
implements our proposed model of polynomial spectral kernel learning. The example command below computes a degree 2 polynomial on the weather data using 15 signals as training
python main.py --data=weather --training=15 --degree=2 --constrained=on
The real world datasets used in the paper are fmri, weather, and uber. For each dataset, the following training sizes are used in the paper:
fmri
: 21 and 42uber
: 10 and 20weather
: 15 and 30.
The --constrained
parameter can be specified to off
to skip constrained optimization.
baselines.py
computes the performances from the baseline models, a --model
term needs to be specified instead of --degree
and --constrained
:
python baselines.py --data=weather --training=15 --model=standard
Baseline --model
options are:
standard
laplacian
local_averaging
global_filtering
regularized_laplacian
diffusion
1_random_walk
3_random_walk
cosine
.