Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Introducing New Node Prediction in Graph Mining: Predicting All Links from Isolated Nodes with Graph Neural Networks

Damiano Zanardini, Emilio Serrano

TL;DR

This paper defines the problem of new node prediction, a zero-shot out-of-graph all-links prediction task where a brand-new node has no prior connections. It proposes a Graph Neural Network–based architecture consisting of a GNN for node embeddings, an example generator, and an MLP-based binary predictor that scores candidate neighbor sets. Experiments on synthetic graphs and a subset of ogbl-citation2 demonstrate that the approach can predict a large fraction of links for new nodes, with performance influenced by graph type, purity settings, and whether auxiliary node features are present. The work highlights practical implications, including a possible bibliography recommendation use-case, and discusses how noisy training data can improve generalization, while outlining directions for future work such as more graphs and transformer-style architectures.

Abstract

This paper introduces a new problem in the field of graph mining and social network analysis called new node prediction. More technically, the task can be categorized as zero-shot out-of-graph all-links prediction. This challenging problem aims to predict all links from a new, isolated, and unobserved node that was previously disconnected from the graph. Unlike classic approaches to link prediction (including few-shot out-of-graph link prediction), this problem presents two key differences: (1) the new node has no existing links from which to extract patterns for new predictions; and (2) the goal is to predict not just one, but all the links of this new node, or at least a significant part of them. Experiments demonstrate that an architecture based on Deep Graph Neural Networks can learn to solve this challenging problem in a bibliographic citation network.

Introducing New Node Prediction in Graph Mining: Predicting All Links from Isolated Nodes with Graph Neural Networks

TL;DR

This paper defines the problem of new node prediction, a zero-shot out-of-graph all-links prediction task where a brand-new node has no prior connections. It proposes a Graph Neural Network–based architecture consisting of a GNN for node embeddings, an example generator, and an MLP-based binary predictor that scores candidate neighbor sets. Experiments on synthetic graphs and a subset of ogbl-citation2 demonstrate that the approach can predict a large fraction of links for new nodes, with performance influenced by graph type, purity settings, and whether auxiliary node features are present. The work highlights practical implications, including a possible bibliography recommendation use-case, and discusses how noisy training data can improve generalization, while outlining directions for future work such as more graphs and transformer-style architectures.

Abstract

This paper introduces a new problem in the field of graph mining and social network analysis called new node prediction. More technically, the task can be categorized as zero-shot out-of-graph all-links prediction. This challenging problem aims to predict all links from a new, isolated, and unobserved node that was previously disconnected from the graph. Unlike classic approaches to link prediction (including few-shot out-of-graph link prediction), this problem presents two key differences: (1) the new node has no existing links from which to extract patterns for new predictions; and (2) the goal is to predict not just one, but all the links of this new node, or at least a significant part of them. Experiments demonstrate that an architecture based on Deep Graph Neural Networks can learn to solve this challenging problem in a bibliographic citation network.
Paper Structure (17 sections, 2 equations, 2 figures, 7 tables)

This paper contains 17 sections, 2 equations, 2 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Preliminaries
  4. The problem of new node prediction
  5. The Deep Learning architecture for Node Prediction
  6. Experiments
  7. Datasets
  8. Experimental setup
  9. Results
  10. Comparison on different GNN architectures
  11. Comparison on the type of graph
  12. Comparison on graph size
  13. Comparison on the type of node features
  14. Comparison on different values for train and test purity
  15. Comparison on the shape of test examples
  16. ...and 2 more sections

Figures (2)

  • Figure 1: (a) A tiny graph $G$ with 8 nodes; (b) the training portion $G_{\mathit{train}}$ of it: $\{n_1,n_2,n_3,n_4,n_5,n_6\}$ are the training nodes, whereas $\{n_7,n_8\}$ are the test nodes; (c) the graph $G_{\mathit{train}}^-$ of negative edges; (d) some (positive or negative) training examples with $\mathit{minPure} = 0.7$ and $\mathit{maxSpurious} = 0.5$ (target nodes are boxed; spurious edges are dotted).
  • Figure 2: A Deep Learning architecture for the new node prediction problem.

Theorems & Definitions (2)

  • Example 1
  • Example 2