thesis_nmf

This repository collects all the code used in my MSc thesis on matrix factorization techniques for dimensionality reduction and clustering.

Datasets

We used three different datasets for our comparative study.

Key information:

• Synthetic: Located in /synthetic_data, data generation function is present in synthetic_uniform.py.
• Leukemia: Located in /leukemia_data, data has been retrieved from this link.
• Barley: Located in /barley, meta information has been retrieved from this DOI and SNP marker data has been provided by Jonathan Kunst.

Code Organization

genolytics

PCA and (penalized) NMF benchmarking has been implemented in the genolytics package.

Specifically:

• models: Implements the PCCLustering and NMFmodel which allow for easy execution of dimension reduction, followed by k-means clustering.
• sqlalchemy: Used to keep track of different runs.
• genolytics.benchmark: Implements the logic to execute a parameterized run and write the result into a SQLite database.
• genolytics.executions: Defines and executes the desired benchmarking tests.

As the Bayesian NMF models have been executed in R, they are not part of the genolytics package. However, genolytics.bnmf_dive collects scripts used to analyze the Bayesian results. The Bayesian NMF scripts can be found in examples.

examples

For convenient try-outs, I added example scripts for each model in examples. Instead of executing the whole benchmarking procedure, single runs will be executed. The scripts are already pre-parameterized, but can be easily manipulated to test different cases.