Extract the Knowledge of Graph Neural Networks and Go Beyond it: An Effective Knowledge Distillation Framework

2021 ◽

Author(s):

Xiang Deng ◽

Zhongfei Zhang

Keyword(s):

Neural Networks ◽

Back Propagation ◽

Multinomial Distribution ◽

Large Data ◽

Training Data ◽

Grid Data ◽

Continuous Space ◽

Graph Data ◽

Knowledge Distillation ◽

Graph Neural Networks

Knowledge distillation (KD) transfers knowledge from a teacher network to a student by enforcing the student to mimic the outputs of the pretrained teacher on training data. However, data samples are not always accessible in many cases due to large data sizes, privacy, or confidentiality. Many efforts have been made on addressing this problem for convolutional neural networks (CNNs) whose inputs lie in a grid domain within a continuous space such as images and videos, but largely overlook graph neural networks (GNNs) that handle non-grid data with different topology structures within a discrete space. The inherent differences between their inputs make these CNN-based approaches not applicable to GNNs. In this paper, we propose to our best knowledge the first dedicated approach to distilling knowledge from a GNN without graph data. The proposed graph-free KD (GFKD) learns graph topology structures for knowledge transfer by modeling them with multinomial distribution. We then introduce a gradient estimator to optimize this framework. Essentially, the gradients w.r.t. graph structures are obtained by only using GNN forward-propagation without back-propagation, which means that GFKD is compatible with modern GNN libraries such as DGL and Geometric. Moreover, we provide the strategies for handling different types of prior knowledge in the graph data or the GNNs. Extensive experiments demonstrate that GFKD achieves the state-of-the-art performance for distilling knowledge from GNNs without training data.

Download Full-text

Graph Neural Networks for Prediction of Fuel Ignition Quality

10.26434/chemrxiv.12280325.v1 ◽

2020 ◽

Author(s):

Artur Schweidtmann ◽

Jan Rittig ◽

Andrea König ◽

Martin Grohe ◽

Alexander Mitsos ◽

...

Keyword(s):

Neural Networks ◽

Octane Number ◽

Molecular Graph ◽

Chemical Properties ◽

Graph Representation ◽

Structure Property ◽

Oxygenated Hydrocarbons ◽

Physico Chemical ◽

Ignition Quality ◽

Graph Neural Networks

<div>Prediction of combustion-related properties of (oxygenated) hydrocarbons is an important and challenging task for which quantitative structure-property relationship (QSPR) models are frequently employed. Recently, a machine learning method, graph neural networks (GNNs), has shown promising results for the prediction of structure-property relationships. GNNs utilize a graph representation of molecules, where atoms correspond to nodes and bonds to edges containing information about the molecular structure. More specifically, GNNs learn physico-chemical properties as a function of the molecular graph in a supervised learning setup using a backpropagation algorithm. This end-to-end learning approach eliminates the need for selection of molecular descriptors or structural groups, as it learns optimal fingerprints through graph convolutions and maps the fingerprints to the physico-chemical properties by deep learning. We develop GNN models for predicting three fuel ignition quality indicators, i.e., the derived cetane number (DCN), the research octane number (RON), and the motor octane number (MON), of oxygenated and non-oxygenated hydrocarbons. In light of limited experimental data in the order of hundreds, we propose a combination of multi-task learning, transfer learning, and ensemble learning. The results show competitive performance of the proposed GNN approach compared to state-of-the-art QSPR models making it a promising field for future research. The prediction tool is available via a web front-end at www.avt.rwth-aachen.de/gnn.</div>

Download Full-text