Social Robot Detection Based on LLM and LSTM-LDA Model
Sheng Wang, Jingyuan Li, Shiqi Sun, Bo Li, Qiong Wu, Haoliang Zhang · 2024
In recent years, the development of short video platforms has been very rapid, and the number of its users is also growing rapidly, unlike traditional social platforms, short video platforms have attracted a large number of users to join due to their accurate push function and unique interaction methods, and play an important role in social platforms. On the other hand, a large number of social robots have also appeared in the short video platform, which are used to spread false information, rumors or misleading content, which has had a negative impact. In order to better detect these social bots, we collected data of about 500,000 users from Kuaishou, the top 2 short video platforms, and extracted four characteristics that are unique to social bots on short video platforms: abnormal daily publishing frequency, abnormal posting interval, abnormal similarity of posting and comment content, and abnormal number of activities. We use LSTM and LDA models to capture the emotional tendencies of user reviews and the topical similarity between the original video content and its comments, which helps to identify social bots more accurately. At the same time, we also use a large language model(LLM) to verify the emotional tendencies obtained to ensure accuracy. Compared with other detection results obtained using only random forest models, our proposed method for social robot detection on short video platforms based on LSTM and LDA models shows an increase in recall rate of about 4% and an increase in F1 score of about 4%.