The identification of diffusion sources in a network is an interesting and important problem. A solution to this problem may be used, for example, for solving security problems such as finding the source of false or malicious information being spread through a network.
When designing a solution for a problem such as this one that has real-world applications, it is important to keep in mind that solutions should be realistic. This is one area where much of the previous work on identification of diffusion sources falls short. For example, much of the previous work assumes that snapshots of the entire network are available. In this paper, the authors propose five technical challenges that reflect common real-world conditions. They then propose a learning model that accurately detects diffusion sources under these conditions.
Both a theoretical analysis of the model and experimental validation are provided. Four of the datasets used for the experiments are synthetic, but one is based on a crawl of a real social network, further emphasizing the real-world applicability of these results. The results show that the proposed model outperforms benchmark models (meaning the predicted diffusion sources are closer to the actual diffusion sources) on a variety of datasets and conditions.