PiNet: Attention Pooling for Graph Classification

Work on this project is ongoing and is described in detail in the original paper.

If you use this work, please cite the following paper:

Meltzer, P., Mallea, M. D. G., & Bentley, P. J. (2019). PiNet: Attention Pooling for Graph Classification, NeurIPS 2019 Graph Representation Learning Workshop

If this work is of interest to you, or you use the code, I'd love your feedback. Please email me at p.meltzer@cs.ucl.ac.uk with any comments, criticisms or suggestions! :D

A Pytorch implementation of Pinet can be found here.

Example Usage

Building The Model

out_dim_a2: output dimension for attention
out_dim_x2: output dimension for features
learn_pqr: learn message passing parametrisation during training (default is True)
preprocess_A: List of options as Strings (for manual pre-processing - should not be used with learn_pqr=True)
- 'add_self_loops'
- 'sym_normalise_A'
- 'laplacian'
- 'sym_norm_laplacian'
- could include multiple, i.e.: ['add_self_loops', 'sym_normalise_A']
tensor_board_logging: enable logging for TensorBoard
reduce_lr_callback: reduce learning rate based on validation set

from model.PiNet import PiNet
from analysis.experiment2 import generate
from sklearn.model_selection import StratifiedKFold

num_classes = 3
batch_size = 5

pinet = PiNet()

folds = list(StratifiedKFold(n_splits=10, shuffle=True).split(X, Y))

Data Format

A: List of adjacency matrices as ndarrays
X: List of features matrices as ndarrays
Y: (n x 1) ndarray containing class no.

Evaluation

num_classes: number of classes
epochs: default 200
batch_size: default 50
folds: Folds or splits of train/test ids
dataset_name: default is 'dataset_name'
verbose: default is 1

accs, times = pinet.fit_eval(A, X, Y, num_classes=num_classes,
                                  epochs=50, batch_size=batch_size, folds=folds, verbose=0)

Prediction only

preds = pinet.get_predictions(A, X, Y, batch_size=batch_size)

GCNLayer

GCN: model/MyGCN.py
MyGCN layer takes a list of A_ and X^{l} as input, and gives a single output of X^{l+1}

Data Generator

analysis/experiment2/generate.py

Experiments

Experiment 1: Isomorphism

params:

num_nodes_per_graph = 50
num_graph_classes = 5
num_node_classes = 2
num_graphs_per_class = 100
batch_size = 5
examples_per_classes = [2, 4, 6, 8, 10]

train set selected by stratified sample
repeated 10x per examples_per_classes

Experiment 2: Message Passing Mechanisms

Observe effect of various matrices for message passing/diffusion.

Experiment 3: Benchmark Against Existing Methods

Compare against existing methods on benchmark data.

Mean classification accuracies for each classifier. For manual search the values p and q as follows: MUTAG and PROTEINS p = 1, q = 0, NCI-1 and NCI-109 p = q = 1, PTC p = q = 0. * indicates PiNet (both models) achieved statistically significant gain.

meltzerpete/PiNet