GADEC: Discovering Abnormal Citation Groups Based on Enhanced Local Community Expansion and DQN

Scholars are enhancing their academic publication for research sharing, professional development, opportunities for collaboration, and so on. They will win academic prestige if their publications have a high citation index due to the excellent quality. However, some scholars ally with each other to build a group (community) for over-citing each other’s papers, in order to increase citation amount fraudulently and further angle for praises and benefits. Such a group demonstrates a tightly linked citation relationship in academic social network (graph). Many existing anomaly group detection methods don’t extract and utilize rich semantic information from network and neglect the joint optimization of dense subgraphs discovery and node representation learning, thus the detection performance is degraded. To deal with these issues, an anomalous citation group detection approach GADEC is proposed, which bases on local expansion community discovery and DQN (Deep Q Network). Semantic information such as research field, paper title, and self-citation is extracted and represented to richen authors’ abnormal features. An expansion measure function is enhanced for community discovery, which integrates several key metrics like node transfer similarity, node community membership, abnormal citation degree, etc. GADEC can accurately find the anomaly citation groups by mutually promoting both detection performance of anomalous communities and authors. The experiments on real data sets have demonstrated very good effectiveness of GADEC on the abnormal citation group mining.