Probabilistic Logic Graph Attention Networks for Reasoning

Knowledge base completion, which involves the prediction of missing relations between entities in a knowledge graph, has been an active area of research. Markov logic networks, which combine probabilistic graphical models and first order logic, have proven to be effective on knowledge graph tasks like link prediction and question answering. However, their intractable inference limits their scalability and wider applicability across various tasks. In recent times, graph attention neural networks, which capture features of neighbouring entities, have achieved superior results on highly complex graph problems like node classification and link prediction. Combining the best of both worlds, we propose Probabilistic Logic Graph Attention Network (pGAT) for reasoning. In the proposed model, the joint distribution of all possible triplets defined by a Markov logic network is optimized with a variational EM algorithm. This helps us to efficiently combine first-order logic and graph attention networks. With the goal of establishing strong baselines for future research on link prediction, we evaluate our model on various standard link prediction benchmarks, and obtain competitive results.