radajakub/gml-dim-reduction

Dimensionality reduction using graphs

GMLFinalProject

Package to develop new dimensionality reduction techniques using graphs and graph machine learning.

Installation

To run the final_notebook_colab.ipynb on Google Colab no special installation is necessary. The notebook will prompt the user to upload source.zip file which contains our codes from ./src/ folder and is located at the same level as final_notebook_colab.ipynb. This is the easiest way to use this package in Colab.

To run the package and final_notebook.ipynb locally it is required to install dependencies by using:

pip install --upgrade pip
pip install -r requirements.txt

Using virtual environment with python 3.6 is recommended.

Almost all functionality is enabled by default, however, python 3.6 is required for running the GraphSAGE embedding algorithm. This algorithm will not work otherwise. The StellarGraph library requires python 3.6, so using this specific version is highly recommended.

This code does not use GPU.

Usage

Example usage of the provided functions and classes is showcased in final_notebook.ipynb.

The structure of the project consists of Builders and Embedders.

• Builders construct graphs of different types from input data points.
  • weight and feature functions can be chosen from ./src/utils/weights.py and ./src/utils/features.py respectively
  • note that only GraphSAGE uses and requires node features, weights are used by all algorithms
• Embedders takes a graph built by a Builder and embeds it into a space of specified dimension.

More Builders than showcased in the final_notebook.ipynb are in ./src/utils/build.py and Embedders are in ./src/utils/ebmedding.py.

These two instances are then passed into a function reduce_dimension(data, builder, embedder) from ./src/utils/dim_reduction.py which connects them and returns the embedding of the input data.

To visualize graphs and final embeddings use functions provided in ./src/utils/visualization.py, to measure quality of the embeddings use function print_evaluation(data, embeddings) in ./src/utils/evaluation.py.

Example

In this section we show a simple example usage of this package on the Swiss roll dataset.

First we import all necessary packages and modules.

from sklearn.datasets import make_swiss_roll
from utils import embedding, build, visualization, weights, features, dim_reduction, evaluation

Then, the dataset needs to be generated. It can also be displayed.

data, labels = make_swiss_roll(n_samples=1000, noise=0.0, random_state=0)
visualization.show_data(data, labels=labels, square=True)

To embed the data a builder and embedder need to be created.

builder = build.CheapestBuilder()
embedder = embedding.KamadaKawaiEmbedder(embedding_dim=2)

Then, the dimensionality reduction is performed and the trustworthiness metric is computed.

embeddings = dim_reduction.reduce_dimension(data, builder, embedder)
evaluation.print_evaluation(data, embeddings)

To show unwrapped swiss roll use tools from visualization module.

visualization.show_data(embedder.embeddings, labels=labels, square=True)