Gaussian Process Regression and Classification on Graphs
The code is loosely inspired on the code from Chapter 18 of Machine Learning: An Algorithmic Perspective (2nd Edition) by Stephen Marsland and implements the algorithms in Chapters 2 and 3 of Gaussian Processes for Machine Learning by C.E. Rasmussen.
GPnet at the moment exposes two classes: GPnetRegressor and GPnetClassifier
- GPnetRegressor provides basic regression functionality for functions defined on graph nodes
- GPnetClassfier provides classification of test nodes, given a set of -1/+1 labels for the training nodes
Available kernels are based on graph node kernels by Kondor et al.
update: kernel composition is already possible via custom function definitions: "soft" kernel composition is yet to be implemented.
Parameter optimization is provided by scipy.optimize.
GPnet demos require networkx, pandas , numpy, matplotlib, scipy.optimize, and random to work: nothing too exotic.
Possible future work directions include:
- including support for multiple and custom-defined kernels
- fixing numerical issues with covariance matrix estimation and parameter optimization