Temporal Interaction Embedding for Link Prediction in Global News Event Graph

Jing Yang

Abstract: Global news events graphs (GNEG) are designed for the noisy and ungrammatical world’s news media, aiming at capturing the true insight and providing explanations by incor- porating potential dimensions and network structures of global news. This paper focuses on the temporal representation learning of GNEG to eliminate misunderstanding or ambiguity caused by missing information. Although some temporal models have been developed, the crossover interactions among entity, relation, and time have not been explicitly discussed. The multi-directional effects between entities, relations, and timestamps matter in predicting the establishment of quadruples. This motivates the proposal of learning temporal interaction embeddings (TIE) to benefit GNEG link prediction performance. Specifically, 1) We propose a crossover convolution layer to learn the two-by-two and common interaction features of entity, relation, and time in GNEG to capture their potential effect patterns in the context of different quadruples; 2)For the learned interaction information, we adopt tensor neural network to maintain the multiple order structure and further extract effective features to improve pre- diction; 3) A tensor temporal consistency constraint is proposed to enhance the learning of time-weakly sensitive information and induce the embeddings to have a certain compatibility over time. Finally, we carried out extensive experiments on three benchmark datasets, the results proved that the performance of the proposed temporal interaction embedding model is competitive with the state-of-the-art methods.