Deep User Profile Construction and Behavior Prediction Based on Multimodal Heterogeneous Feature Fusion

As the platform for users to express their personal opinions and thoughts, Online Social Networks (OSNs) serve as a primary channel for most individuals to receive social information. During social emergencies, the OSNs experience a surge in discussions, some of which may be malicious and negatively affect public opinion. In this paper, we manually construct a dataset of malicious users based on Twitter, named Malicious_Users_2023 (MU2023). Beyond the tweet features, we summarize three sets of supplementary features for malicious user monitoring: the User Feature, the Network Feature, and the High-dimensional Feature. Leveraging all sets of features, a scheme for malicious individuals monitoring utilizing multimodal feature fusion analysis is proposed. Experimental results show that the proposed scheme achieves 84
^%
accuracy on the MU2023 dataset and 96% on the Apontador dataset. Furthermore, we validate the performance enhancement of monitoring with three additional sets of features and two feature fusion approaches, respectively. This is crucial for comprehending and modeling individuals' malicious behavior effectively.