Abstract
The spread of fake news may disturb the financial markets in the background of program trading. Nowadays, news often contains different types of data and circulate on various new media platforms, which challenges to existing fake news detection models. It is not just because the features of fake news are getting more complicated, but also traditional methods often ignore that the user comments may also provide valuable information for this task. To address these gaps, we present a multivariate data feature mining based multilayer framework (MDFM-MF). In the bottom layer, different news carriers are uniformly converted into text data. The middle layer then extracts the opinion, sentiment, user and propagation features from the text data. Eventually, fake news detection is realized based on the BERT model in the top layer. The experiment results show that our proposed model has good performance on detection task.
Introduction
In the high-stakes world of financial markets, a single piece of fake news can trigger algorithmic trading chaos and devastate investor interests. Modern misinformation is a "multivariate" problem, spreading through short videos, inflammatory memes, and audio clips across social media. Traditional detection models often suffer from data silos—focusing on just one medium—and frequently overlook the "wisdom of the crowd" found in user comments. This research introduces the MDFM-MF (Multivariate Data Feature Mining based Multilayer Framework) to bridge these gaps by creating a system that integrates multimodal data and social context.
The MDFM-MF Framework
The framework operates in three distinct stages to process and classify potential misinformation.
The Bottom Layer: Data Harmonization
To handle the complexity of multimedia, the system converts diverse carriers (video, audio, images) into a unified text format. It utilizes Optical Character Recognition (OCR) via Connectionist Text Proposal Network (CTPN) and Convolutional Recurrent Neural Network (CRNN) models to extract text from visual media. This ensures that regardless of the original format, all data can be analyzed using natural language processing techniques.
The Middle Layer: Feature Mining
This layer extracts four critical types of features from the harmonized text to identify the "signature" of fake news:
- Opinion Features: Uses the Enhanced Sequential Inference Model (ESIM) to determine if comments support or oppose the news.
- Sentiment Features: Analyzes emotional category, intensity, and inflammatory language, as fake news often uses high-intensity negative emotions to provoke readers.
- User Features: Introduces a novel User Reliability Index (URI) to weigh input based on a user's historical accuracy in identifying real vs. fake news.
- Propagation Features: Tracks how news spreads through reliable user networks.
The Top Layer: Classification
The extracted features are fed into a BERT (Bidirectional Encoder Representations from Transformers) model for final classification into "Real" or "Fake."
Experimental Evaluation
The framework was evaluated using the Chinese Rumor Dataset (CED_Dataset) and a Self-Built Multivariate Dataset containing 568 videos, 1,034 images, and 190 audio clips.
The results demonstrate that multimodal data is essential; accuracy dropped significantly when relying only on text. By filtering out "noisy" comments and focusing on effective social feedback, the full framework achieved:
- Maximum Accuracy: 90.44%
- Recall: 93.77%
- F1 Score: 0.9274
Conclusion
The MDFM-MF framework successfully leverages multimodal data and user social features to detect financial misinformation. By converting diverse media into a unified text format and mining deep social features like user reliability, the model provides a robust defense against market-destabilizing rumors. Future work will focus on reducing computational overhead for real-time applications and adapting the model to diverse languages and sophisticated adversarial strategies like deepfakes.
Summary updated