kernelbiome 0.1.1

A kernel-based nonparametric regression and classication framework for compositional data.

KernelBiome

Note: This repository contains code that can reproduce outputs in the paper Supervised Learning and Model Analysis with Compositional Data (Huang et al., 2022). The KernelBiome python package can be installed via

pip install kernelbiome

or

python -m pip install git+https://github.com/shimenghuang/KernelBiome.git

Minimum usage example for testing out installation:

import numpy as np
from kernelbiome.kernels_jax import *

x = np.random.uniform(0, 1, 5)
x /= x.sum()
y = np.random.uniform(0, 1, 5)
y /= y.sum()

k_linear(x,y)

The KernelBiome Package

The package contains the following modules:

• metrics_jax.py: distance metrics with jax.
• kernels_jax.py: kernel functions with jax.
• helpers_jax.py: helper functions for kernels and metrics with jax.
• weighted_kernels_jax.py: weighted kernel functions with jax.
• helpers_weighting.py: helpers for calculating weight matrices.
• cfi_and_cpd.py: CFI and CPD calculation.
• nested_cv: functions for nested CV.
• utils_cv.py: utility functions including utilities for nested CV
• utils_result.py: utility functions for result summary.

Reproducible Code

data

This folder should contain two subfolders:

• MLRepo which should contain a subfolder qin2014 containing data directly taken from here.
• CentralParkSoil should contain data pre-processed by prep_centralparksoil.R based on data from here.

scripts

This folder contains scripts that can reproduce resutls in the paper:

• tree_utils.py: utility functions regardign the UniFrac distance.
• load_<ds>.py: load the dataset into proper format.
• create_unifrac_weights_<ds>.py: create the UniFrac based weight matrices.
• setup_params.py: set up kernel parameters and hyperparameters.
• run_cfi_cpd_<ds>.py: calculate CFI and CPD.
• plot_<ds>_cfi.py: make CFI plot.
• plot_<ds>_cpd.py: make CPD plot.
• plot_mds_cirrhotic.py: make MDS plot.
• save_cv_indices_<ds>.py: save the CV indices for the 50-fold CV comparison so that each approach is run on the same part of the data in run_compare_one_fold_<ds>.py.
• run_compare_one_fold_<ds>.py: run one of the 50-fold CV using different methods.

where <ds> is one of cirrhotic or centralparksoil.

notebooks

This folder contains notebooks for demonstration:

• simplex_heatmap.py: heatmap functons to visualize kernels on 2-simplex.
• workflow_demo.ipynb: A workflow demonstration using simulated data.
• visualization_writeup.ipynb: Visualization of kernels with heatmap on 2-simplex.