kstoneriv3 / dpp-nystrom-gp

Gauusian Process with DPP-Nystrom. Carried out as a semester project.

Gauusian Process with DPP-Nystrom

This repository includes the source code for Gaussian process regression combined with the Nystrom approxomation by k-DPP sampling. The results of the numerical experiments can be checked in the ./demo.ipynb or in this report.

Demo on Gaussian Process with Nystrom Method

The following four landmark selection scheme (of Nystrom method) were compared;

• uniform sampling
• greedy algorithm for the likelihood maximization
• k-DPP (by Gibbs sampling)
• simmulated annealing of the MAP of k-DPP

The result can be replicated by running the jupyter notebook contained in the main directory. The dataset used here (aileron dataset) is taken from https://sci2s.ugr.es/keel/dataset.php?cod=93.

---

Prerequisites

• Python 3.7
• numpy
• pandas
• scipy
• matplotlib
• multiprocessing
• time

Folders/Files

.
├──sampler
│   ├─ __init__.py
│   ├─ dpp.py
│   ├─ greedy.py
│   ├─ mcdpp.py
│   ├─ quadrature.py
│   ├─ quadrature_back.py
│   ├─ sadpp.py
│   └─ utils.py
├──helper
│   ├─ __init__.py
│   └─ helper.py
├──data
│   └─ ailerons.txt
├──fig
│   └─ summary.png
├─ demo.ipynb
└─ README.md