Content Sharing Prediction for Device-to-Device (D2D)-based Offline Mobile Social Networks by Network Representation Learning

摘要

With the explosion of cellular data, the content sharing in proximity among offline Mobile Social Networks (MSNs) has received significant attention. It is necessary to understand the face-to-face (e.g. Device-to-Device, D2D) social network structure and to predict content propagation precisely, which can be conducted by learning the low-dimensional embedding of the network nodes, called Network Representation Learning (NRL). However, most existing NRL models consider each edge as a binary or continuous value, neglecting rich information between nodes. Besides, many traditional models are almost based on small-scale datasets or online Internet services, severely confining their applications in D2D scenarios. Therefore, we propose ResNel, a RESCAL-based network representation learning model, which aims to regard the multi-dimensional relations as a probability in third-order (3D) tensor space and achieve more accurate predictions for both discovered and undiscovered relations in the D2D social network. Specifically, we consider the Global Positioning System (GPS) information as a critical relation slice to avoid the loss of potential information. Experiments on a realistic large-scale D2D dataset corroborate the advantages of improving forecast accuracy.