New Model Enhances Dynamic Knowledge Graph Link Prediction

Dynamic knowledge graphs are fundamental to many AI applications, representing complex relationships that evolve over time, such as molecular structures or social networks. A key challenge is accurately modeling these temporal and relational dependencies, especially when data is noisy or incomplete. This research proposes a new probabilistic model called PGRE (Poisson-Gamma Relational Evolution) to address these issues. PGRE employs a Poisson-Bernoulli framework to represent multi-relational temporal links. It introduces Gamma-distributed latent variables to capture how entities associate with factors and how different relations depend on each other through shared latent communities. Furthermore, a Gamma Markov process models the temporal evolution of these latent variables, providing a principled way to characterize relational dynamics. Experiments show PGRE performs competitively in link prediction, particularly in sparse datasets, while also revealing meaningful patterns in how relationships change over time.