Graph classification using kernel methods

This repository is the work produced by Ludovic De Matteïs and Matias Etcheverry in the course Machine learning with kernel methods given by Michael Arbel, Alessandro Rudi, Jean-Philippe Vert and Julien Mairal at the MVA.

Objective

The goal of this repository is to implement machine learning algorithms, for a classification task on graph data.

Installation

In order to run this repository, you need to run the following:

# clone the repository
git clone git@github.com:MatiasEtcheve/KM-graph-classification.git
cd KM-graph-classification

# install the dependencies
pip install -r requirements.txt

Inference

The Command Line Interface can be used efficiently to compute predictions on the dataset.

In order to use our best model, you can run:

python start.py

This will train 2 kernels:

• a Weisfeiler Lehman kernel on the edges of the graphs
• a Weisfeiler Lehman kernel on the nodes of the graphs A linear combination is then applied between the logits of the 2 models.

However, you can also tune the learning, with the CLI options:

Option name	Type	Description	Default Value
kernels	list	Kernels to train on the data. If multiple kernels are provided, it will individually train on each kernel, then do a linear combination of the logits. Must be one (or multiples) of "EH", "VH", "SP", "GL", "WL-Edges", "WL-Nodes"	"[WL-Edges,WL-Nodes]" (care quotes !)
combination	list	list of coefficient for the combination of kernels	"[1.59,1.35]" (care quotes !)
max-alpha	float	Max value of the alpha coefficient in SVM. Note: multiple alphas can be higher than C, when class_weight=balanced.	100
sigma	float >= 0	Sigma in the RBF wrapper. If 0, a linear wrapper is applied instead.	Defaults to 1
src	folder	Path to .pkl datasets.	data/
train-val-split	float or int	Train val split, in ratio or in number of elements, eg 0.7 or 4200. Usefull when training takes time.	0.7
do-predict	flag	whether to do the prediction on the test set.	True
predict-filename	filename	path to the prediction (if do-predict flag is present).	test_pred.csv

Example: Examples of working command lines:

• Default command: python start.py --kernels "[WL-Edges,WL-Nodes]" --combination "[1.59,1.35]" --src data/ --train-val-split 0.7 --do-predict --predict-filename data/predictions.csv
• another command: python start.py --kernels "[EH,VH]" --combination "[1,1]" --max-alpha 1 --sigma 0 --src data/ --train-val-split 0.4 --do-predict --predict-filename data/predictions.csv