Research On Key Technologies Of Event Prediction Based On Graph Representation Learning

With the development of modern society,computer technology and artificial intelligence,more and more organizations want to use the capabilities of artificial intelligence to predict possible future events.Event prediction is a subject of great social and economic value.How to mine the development trajectory,the development law and the development mode of events are very important.This research focuses on graph representation learning based methods for event prediction.This research first explores the event prediction based on homogeneous graph representation learning,focusing on the script event prediction task.The development law of events and the connection relationship between events are intricate,and it is difficult to achieve a better effect in event prediction tasks by only modeling partial information such as event pairs or event chains.To this end,this article builds an eventic graph to model the development of events,and uses a framework that combines a pre-trained language model and a graph neural network to predict subsequent event.The pre-trained language model encodes events in a deep way,graph neural network models structure information of graph.Second,this research explores the event prediction based on heterogeneous graph representation learning,we focus on the task of stock market prediction.Previous researches on stock market prediction only used a single semantic unit for forecasting,and it is difficult to accurately capture key information and new information in financial text.For this reason,this paper proposes an event prediction framework based on heterogeneous graph representation learning.It constructs a heterogeneous graph by using information of different granularities in financial text.The coarse-grained information supplements the global information,and the finegrained information can refine the coarse-grained information for fine-grained information selection,sequential heterogeneous graph neural networks are used to exchange and update information among different-grained nodes.Finally,after trying the event prediction based on homogeneous graph and heterogeneous graph representation learning,this research explores event prediction in the financial field.According to the idea of homogeneous graph representation learning,this article first expands the financial eventic graph.Eventic graph can provide evidence events as an external knowledge base,and then we propose a financial risk event prediction framework enhanced by eventic knowledge.By introducing hyperbolic space,the structural distortion of graph represented in Euclidean space is reduced.And the directionality of causality is captured with our proposed graph neural networks,so that the information of evidence events can be transmitted more reasonably.Frameworks proposed in this study have achieved higher performance in event prediction tasks such as script event prediction,stock market prediction,and causal reasoning.The experimental results and ablation study also prove the feasibility and effectiveness of our models.