Event prediction through structural intelligence in online social networks

Event prediction through structural intelligence in online social networks

The Internet today is a platform of information exchange between real people across the globe. Event prediction is an emerging and highly complex topic of interest which enjoys wide ranging applications in fintech, medical, security, etc. Some of these implementations include time sequenced methods, pattern recognition techniques, multiple instance learning, topic based approach, etc. While they have been adequate at handling predictions of events from past discrete occurances, they fall short of the capability to predict events from a continuous stream of social information exchange.

Furthermore, many of these approaches lack the presentative power of describing and tracking events through time. Relational flux and turbulence in Online Social Networks (OSNs) can be defined as the complex evolution of social communication patterns staged over important topic contexts which have the potential to cause abberations of relational states. They play very important roles in determining tasks like recognition, prediction, detection, etc. across applications like recommendation, clustering, community, privacy, security, knowledge representation, etc. For example, an essential research question for Knowledge Representation Learning (KRL) is how to explictly embed implict real-life relational states between entities structured in a Knowledge Graph (KG).

Most current studies today however, do not have the capability to effectively generalize relationships across heterogeneous architectures. Indeed, an important challenge to address is that latent communication patterns in local and global contexts of social opinions cannot be fully captured. Thus, event prediction is challenging for two reasons: its generalized, temporal, evolving nature and drifting contexts. In addition, many current approaches however, lack the capacity of describing and tracking general events over time. To tackle these issues, this study develops a novel RFT model which leverages on the mechanics of Relational Flux and Turbulence to model dynamic communicative behaviors between actors within social networks. To the best knowledge offered by existing literature, there has not been a similar model and / or method of approach which effectively predicts events from a computationally cognitive perspective.

To surmise the milestones achieved by this research endeavour, extensive experiments were conducted on large-scale datasets from Twitter, Googlefeed and Livejournal. From the experimental results, it was shown that RFT is able to identify and predict relational turbulence in a social flux which mirrors real life relational state transitions in a social topic context. The following demonstration from the F1-scores and k-fold cross validation results proves that the model performs comparably better by more than 10% to well-known predictors such as the Hybrid Probabilistic Markovian (HPM)
predictive method [1] and other state-of-the-art baselines in predicting events. Importantly, this research development proves that event prediction methods which account for relational features between actors of social networks perform much better than conventional mainstream approaches like vector regression, random walk, markovian logic networks, etc. that are widely used today.