Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Link-aware link prediction over temporal graph by pattern recognition

Bingqing Liu, Xikun Huang

TL;DR

The paper addresses temporal-link prediction by moving beyond node-representation learning to a link-aware, pattern-recognition framework. It introduces TGACN, which jointly processes the query link and recalled historical links using parametric and nearest sampling, dual attention to capture patterns at multiple granularities, and a CNN-based pattern recognizer with CAM-based interpretability. The method achieves state-of-the-art performance on six datasets and offers transparent explanations of which historical links influenced predictions. This pattern-centric approach advances practical temporal link prediction and suggests future extensions to static graphs and broader interpretability.

Abstract

A temporal graph can be considered as a stream of links, each of which represents an interaction between two nodes at a certain time. On temporal graphs, link prediction is a common task, which aims to answer whether the query link is true or not. To do this task, previous methods usually focus on the learning of representations of the two nodes in the query link. We point out that the learned representation by their models may encode too much information with side effects for link prediction because they have not utilized the information of the query link, i.e., they are link-unaware. Based on this observation, we propose a link-aware model: historical links and the query link are input together into the following model layers to distinguish whether this input implies a reasonable pattern that ends with the query link. During this process, we focus on the modeling of link evolution patterns rather than node representations. Experiments on six datasets show that our model achieves strong performances compared with state-of-the-art baselines, and the results of link prediction are interpretable. The code and datasets are available on the project website: https://github.com/lbq8942/TGACN.

Link-aware link prediction over temporal graph by pattern recognition

TL;DR

The paper addresses temporal-link prediction by moving beyond node-representation learning to a link-aware, pattern-recognition framework. It introduces TGACN, which jointly processes the query link and recalled historical links using parametric and nearest sampling, dual attention to capture patterns at multiple granularities, and a CNN-based pattern recognizer with CAM-based interpretability. The method achieves state-of-the-art performance on six datasets and offers transparent explanations of which historical links influenced predictions. This pattern-centric approach advances practical temporal link prediction and suggests future extensions to static graphs and broader interpretability.

Abstract

A temporal graph can be considered as a stream of links, each of which represents an interaction between two nodes at a certain time. On temporal graphs, link prediction is a common task, which aims to answer whether the query link is true or not. To do this task, previous methods usually focus on the learning of representations of the two nodes in the query link. We point out that the learned representation by their models may encode too much information with side effects for link prediction because they have not utilized the information of the query link, i.e., they are link-unaware. Based on this observation, we propose a link-aware model: historical links and the query link are input together into the following model layers to distinguish whether this input implies a reasonable pattern that ends with the query link. During this process, we focus on the modeling of link evolution patterns rather than node representations. Experiments on six datasets show that our model achieves strong performances compared with state-of-the-art baselines, and the results of link prediction are interpretable. The code and datasets are available on the project website: https://github.com/lbq8942/TGACN.
Paper Structure (12 sections, 9 equations, 4 figures, 7 tables)

This paper contains 12 sections, 9 equations, 4 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. THE TGACN MODEL
  4. Problem Formulation and Notations
  5. Method
  6. Experiment
  7. Experimental Setup
  8. Results
  9. Ablation Study
  10. Hyperparameter Investigation
  11. Interpretability
  12. Conclusion

Figures (4)

  • Figure 1: An example temporal graph. Temporal subgraph with red links shows an example pattern, i.e., {(A,B,$t_1$),(B,A,$t_3$)}, which means that B will interact with A after A interacts with B.
  • Figure 2: Patterns of different granularity. Pattern (a) shows that B will respond to A after A interacts with B, pattern (b) describes that no matter whom A interacts with, afterwards this person will response to A, whereas pattern (c) reveals that such sequence holds for any two arbitrary nodes.
  • Figure 3: The overall pipeline of the proposed Temporal Graph Attention Convolution Network (TGACN). Given the query link (the link from B to A with purple color), TGACN first uses nearest sampling and parametric sampling to recall historical links that may help to predict the query link. Then we encode the input links by two kinds of attention: transductive attention and inductive attention, where node identities information and the inductive structural information are preserved. Finally, a convolutional neural network EffNet is utilized to recognize whether the input implies a reasonable pattern that ends with the query link.
  • Figure 4: Illustration of our model's interpretability. On the left of this Figure, we present three ground truth query links and sample four historical links for each (the end of input is the query link). On the right, by CAM, we show the importance of four historical links to the query link for each input.