Dual-Task Network Embeddings for Influence Prediction in Social Internet of Things

Abstract

Social Internet of Things (SIoT) is an emerging application area that supports the spread of multimedia information. Detecting critical nodes based on the results of influence prediction is the main solution to maximizing the information propagation in SIoT. However, while the tasks of predicting an influence probability and a cascade size are both critical in influence prediction, there has not yet a study that thoroughly investigates these two influence parameters. This article presents an end-to-end method that learns dual-task network embeddings to jointly predict influence probabilities and cascade sizes, which is called a multidimensional influence-to-vector method (Multi-Influor). First, multidimensional influence contexts are generated based on random walks, incorporating multiple estimating factors for pairwise node interaction, network structure, and global preference similarity. A new method of learning dual-task network embeddings is then devised to simultaneously capture influence probabilities and cascade sizes. The two tasks are both formulated into a unified framework via enforcing an information-sharing embedding matrix. Finally, stochastic gradient descent (SGD) is used to optimize two loss functions for influence prediction in an alternating manner, and the tasks are jointly accomplished that produces an accurate prediction. Extensive simulation results on real-world data sets show that Multi-Influor outperforms six state-of-the-art methods in accuracy and efficiency, and that a joint training method for the two tasks improves the overall prediction performance. Moreover, Multi-Influor is a practical method for SIoT sustainable computing.