Symbolic Prompt Tuning Completes the App Promotion Graph

Zhongyu Ouyang, Chunhui Zhang, Shifu Hou, Shang Ma, Chaoran Chen, Toby Li, Xusheng Xiao, Chuxu Zhang, Yanfang Ye

Abstract: Recent mobile applications (i.e., apps) have been extensively implanted with paid advertisements that promote other mobile apps, including malware that raises alarming concerns in cybersecurity. Excavating the app promotion patterns in the app-promoting ecosystem allows for early interceptions of malware installment, and hence has gained more attention in recent research. However, related data in the app-promoting ecosystem such as app developers and categories is often scarce, especially when the data is collected from a single data source. The scarce data is insufficient in training effective deep and complex models for app promotion pattern mining, and targeting the data scarcity problem is therefore the key to advancing research in app promotion pattern mining. Therefore, we aim to complete data in the app-promoting ecosystem to pave the way for app-promoting pattern mining. We present SymPrompt, a language model-based framework that leverages the symbolic prompts to complete the missing data in the app-promoting ecosystem. The symbolic prompts are tokens that provide extra contextual information that assists the model in completing the missing data. We devise two sets of symbolic prompts containing contextual information from the perspectives of data structure and data semantics to assist the model prediction. Through extensive experiments, we demonstrate SymPrompt ’s effectiveness in completing the missing in the app-promoting ecosystem. Code: https://github.com/zyouyang/SymPrompt