Summary: | Identifying fraudulent campaigns or messages remains a difficult task in the field of natural language processing. We
proposed a hypothesis that if the well-known brands in the same category as the crowdfunding project can implement similar
technology as crowdfunding projects, the project is considered to be more feasible. The opposite is considered more likely to be
fraudulent. This research proposed a novel BNB-NO-BK method to detect fraudulent crowdfunding projects. A novel method
called BNB, which was constructed by key-BERT, NLTK, and fine-tuned QA model for BERT, was proposed to extract the
characteristics of crowdfunding projects. We proposed a novel NO (Nice Classification & Ontology) method for classifying the
categories of projects, which constructed ontology trees based on the characteristics of the crowdfunding projects and our
modified Nice Classification. Furthermore, we proposed a novel BK (Brand Knowledge) cross-checking method to extract the
features of crowdfunding projects. Finally, we compared the performances of different machine learning methods for identifying
fraudulent crowdfunding projects. Furthermore, to address the problem of possible bias caused by unbalanced data, we used data
augmentation to process the dataset. Our proposed method achieved an accuracy of 95.71% in detecting fraudulent crowdfunding
projects, which was superior to existing methods.
|