An Improved Deep Learning Network, Addressing Graph Node Imbalance in Social Media Rumor Source Detection

Finding the source of rumors in the social network was addressed by researchers with probabilistic models like Maximum Likelihood Estimation in complex network analysis for the past few decades. However, the most promising results could reach up to 2-hop distant neighborhoods on average. With the advent of graph neural networks, the issue was addressed as a classical node classification problem in large networks. Node classification problems achieve the best results when there are appropriate node attributes as features and when node classes are balanced. However, unlike other node classification scenarios, the data collected for source identification usually lacks node attributes because of time limitations. Moreover, the detection of the sources among thousands of users is typically a class imbalance problem. If we could deal with these issues skillfully, then the dominance of the deep learning method compared to other conventional probabilistic methods in multiple rumor source detection will be prominent. We have proposed here a deep learning-based multiple source node classification framework that can predict the sources with promising accuracy. The primary hurdles in non-attributed network classification are navigated by generating feature vectors that capture the structural characteristics of the network and spreading pattern. These features are further solidified with the Graph embedding technique, incorporating the neighborhood features. We have triumphed over the challenge of imbalanced node classes by synthetic node generation with a suitable mathematical model. The concern is further resolved by selective sampling and weighted loss estimation in the deep learning network for classification used in the framework. A study of the Immuno-Diffuse Likelihood parameter in Label propagation based feature construction and its influence on accurate prediction is examined. Our approach demonstrates superior performance compared to existing methods in the datasets available in public repositories, making it a reliable and robust tool for rumor source detection in the complex landscape of social media.