graph-attention-student 0.10.0

Computational Experiments

It is possible to list, show and execute all the computational experiments using a command line interface CLI.

NOTE Most of the experiments have a long runtime, ranging from ~2hrs to ~2days. Furthermore, all of the experiments which do model training are currently configured to run on a GPU and might crash if the GPU can either not be detected or does not have enough VRAM. This setting can be changed in the corresponding experiment scripts

All the available experiments can be listed like this:

python3 -m graph_attention_student.cli list

The details for a specific experiment can be viewed like this:

python3 -m graph_attention_student.cli info [experiment_name]

A new run of an experiment can be started like this. However, be aware that this might take some time.

python3 -m graph_attention_student.cli run [experiment_name]

Each experiment will create a new archive folder, which will contain all the artifacts (such as visual examples and the raw data) created during the runtime. The location of this archive folder can be found from the output generated by the experiment execution.

Archived Experiments

To view the detailed data which was used in the making of the paper, go to graph_attention_student/experiments. The subfolders in that folder contain the archived experiments. These contain extensive examples for each repetition of the various experiments as well as all of the raw data collected during the execution of the experiments.

MEGAN in code

The MEGAN model is implemented as the MultiAttentionStudent class, which implements keras.Model. The implementation is based on the kgcnn library for graph convolutional networks for keras. For further information on loading graph structured data with kgcnn visit: https://github.com/aimat-lab/gcnn_keras

This is a simple example of how to use the model in the regression case:

import tensorflow as tf
import tensorflow.keras as ks
from graph_attention_student.training import NoLoss
from graph_attention_student.models import Megan

model = Megan(
    # These lists define the number of layers and the number of hidden units in each layer for the
    # various parts of the architecture
    units=[9, 9, 9],  # The main convolutional layers
    importance_units=[],  # The MLP that creates the node importances
    final_units=[5, 1],  # The final MLP for graph embeddings
    # Example for a regression problem. We need the prior knowledge about what range the values of the
    # dataset will be expected to fall into...
    regression_limits=(-3, +3),
    # ... as well as a reference value.
    regression_reference=0,
    # This controls the weight of the explanation-only train step (gamma)
    importance_factor=1.0,
    importance_multiplier=5,
    # This is the weight of the sparsity regularization
    sparsity_factor=0.1,
)

# The model output is actually a three tuple: (prediction, node_importances, edge_importances).
# This allows the importances to be trained in a supervised fashion. If we don't want that,
# we can simply supply the NoLoss function instead.
model.compile(
    loss=[ks.losses.MeanSquaredError(), NoLoss(), NoLoss()],
    loss_weights=[1, 1, 1],
    optimizer=ks.optimizers.Adam(0.001)
)

# model.fit() ...

—

RB-Motifs Dataset

This is a synthetic dataset, which basically consists of randomly generated graphs with nodes of different colors. Some of the graphs contain special sub-graph motifs, which are either blue-heavy or red-heavy structures. The blue-heavy sub-graphs contribute a certain negative value to the overall value of the graph, while red-heavy structures contain a certain positive value.

This way, every graph has a certain value associated with it, which is between -3 and 3. The network was trained to predict this value for each graph.

The examples shows from left to right: (1) The ground truth explanations, (2) a baseline MEGAN model trained only on the prediction task, (3) explanation-supervised MEGAN model and (4) GNNExplainer explanations for a basic GCN network. While the baseline MEGAN and GNNExplainer focus only on one of the ground truth motifs, the explanation-supervised MEGAN model correctly finds both.

Water Solubility Dataset

This is the AqSolDB dataset, which consists of ~10000 molecules and measured values for the solubility in water (logS value).

The network was trained to predict the solubility value for each molecule.

Movie Reviews

Originally the MovieReviews dataset is a natural language processing dataset from the ERASER benchmark. The task is to classify the sentiment of ~2000 movie reviews collected from the IMDB database into the classes “positive” and “negative”. This dataset was converted into a graph dataset by considering all words as nodes of a graph and then connecting adjacent words by undirected edges with a sliding window of size 2. Words were converted into numeric feature vectors by using a pre-trained GLOVE model.

Example for a positive review:

Example for a negative review:

Examples show the explanation channel for the “negative” class left and the “positive” class right. Sentences with negative / positive adjectives are appropriately attributed to the corresponding channels.