Graph neural networks (GNNs) are commonly applied to relation prediction tasks, such as recommender systems and knowledge graph completion. In these contexts, pre-training GNNs with a neighborhood reconstruction loss can be a useful strategy to improve performance on downstream classification tasks.

Google

University of Michigan - Ann Arbor

For most current applications, graph neural networks are used in one of three tasks: node classification, graph classification, or relation prediction.

Applications and Loss Functions

Relation prediction (link prediction, graph completion, relation inference) is a machine learning task on graph data where the goal is to infer the missing edges of a graph, when given a set of nodes and an incomplete set of edges between these nodes.

Relation Prediction

2. Graph classification: Similar to node classification, applications on graph-level classification are popular as benchmark tasks. Historically, kernel methods were popular for graph classification, and—as a result—some of the most popular early benchmarks for graph classification were adapted from the kernel literature, such as tasks involving the classification of enzyme properties based on graph-based representations.

Graph neural networks graph classification

Relation Prediction with Graph Neural Networks

Node classification is one of the most common benchmark tasks for graph neural networks (GNNs). The standard approach trains a GNN in a fully-supervised manner: node embeddings are passed through a classifier (e.g., an MLP followed by a softmax) and the model is optimized by minimizing the negative log-likelihood (cross-entropy) loss between predicted and true node labels.

Fully-Supervised Node Classification with Graph Neural Networks

Examples of relation prediction tasks include: recommending content to users in social platforms, predicting drug side-effects, or inferring new facts in a relational databases.

Examples of Relation Prediction Tasks

Like node classification, relation prediction blurs the boundaries of traditional machine learning categories — often being referred to as both supervised and unsupervised — and it requires inductive biases that are specific to the graph domain. In addition, like node classification, there are many variants of relation prediction, including settings where the predictions are made over a single, fixed graph, as well as settings where relations must be predicted across multiple disjoint graphs.

Learn Before

Related