Semi-Supervised Learning combining Graph-based with Feature-based for spam calls detection problems

Tóm tắt báo cáo

Spam over Internet Telephony (SPIT) is becoming more and more serious and has attracted considerable attention from telecom providers due to its huge financial harm and user experience. Current anti-spam systems face two major challenges: data scalability and cheating. In this paper, we present technical solutions to try to solve this challenge. We propose a graph-based method, using a more complex GNN architecture suitable for Spam data type, giving better results than GCN. Then, we propose a semi-supervised learning algorithm that combines cotrain and active learning to improve classification performance. Implemented on a dataset collected from a spam call blocking application called Icaller. The results of the implementation demonstrate the effectiveness of the proposed method.