Social Robot Detection on Short Video Platform Based on Random Forest and LDA Model

Bo Li, Shiqi Sun, Jingyuan Li, Qiong Wu, Yao Zheng, Haoliang Zhang

Abstract: With the rapid development of short video platforms, a large number of social robots have emerged, which are controlled by automated programs and actively post content on various short video platforms. Many of these social robots are used to spread false information, rumors, or misleading content, which negatively impact user experience, content quality, and the social environment on short video platforms. In order to better detect these social robots, we collected about 500k user data from Kuaishou, which is one of the top 2 short video platforms, and extracted four features that are closely related to social robot detection: abnormal daily posting frequency, abnormal posting time intervals, abnormal similarity in post and comment content, and abnormal activity times. We used this collected dataset to train a Random Forest model and then tested it on real user data. Additionally, we use LDA topic modeling technique to calculate the topic similarity between the original video content and its comments, thus helping to identify social robots. Compared to other detection results achieved using only the Random Forest model, our proposed approach based on the Random Forest and LDA model for social robot detection on short video platforms has shown an improvement of approximately 5% in recall rate and about 4% in F1 score.