Event Prediction in Online Social Networks

Abstract

Event prediction is a very important task in numerous applications of interest like fintech, medical, security, etc. However, event prediction is a highly complex task because it is challenging to classify, contains temporally changing themes of discussion and heavy topic drifts. In this research, we present a novel approach which leverages on the RFT framework developed in \cite{tan2020discovering}. This study addresses the challenge of accurately representing relational features in observed complex social communication behavior for the event prediction task; which recent graph learning methodologies are struggling with. The concept here, is to firstly learn the turbulent patterns of relational state transitions between actors preceeding an event and then secondly, to evolve these profiles temporally, in the event prediction process. The event prediction model which leverages on the RFT framework discovers, identifies and adaptively ranks relational turbulence as likelihood predictions of event occurrences. Extensive experiments on large-scale social datasets across important indicator tests for validation, show that the RFT framework performs comparably better by more than 10\% to HPM \cite{amodeo2011hybrid} and other state-of-the-art baselines in event prediction.