Modeling the evolution of temporal knowledge graphs with uncertainty

TL;DR

This paper introduces WGP-NN, a novel graph neural network that uses Gaussian processes to model both the evolution and uncertainty of future events in temporal knowledge graphs, achieving state-of-the-art results.

Contribution

The paper presents a new GNN architecture employing Gaussian processes to jointly model event probabilities and their uncertainties in continuous time, which is novel in tKG forecasting.

Findings

Achieves state-of-the-art performance on benchmark datasets

Models uncertainty in future link predictions effectively

Handles complex temporal and structural dynamics without parameters

Abstract

Forecasting future events is a fundamental challenge for temporal knowledge graphs (tKG). As in real life predicting a mean function is most of the time not sufficient, but the question remains how confident can we be about our prediction? Thus, in this work, we will introduce a novel graph neural network architecture (WGP-NN) employing (weighted) Gaussian processes (GP) to jointly model the temporal evolution of the occurrence probability of events and their time-dependent uncertainty. Especially we employ Gaussian processes to model the uncertainty of future links by their ability to predict predictive variance. This is in contrast to existing works, which are only able to express uncertainties in the learned entity representations. Moreover, WGP-NN can model parameter-free complex temporal and structural dynamics of tKGs in continuous time. We further demonstrate the model's state-of-the-art performance on two real-world benchmark datasets.