Identifying Cyber Threats from Financial Social Media: A Hybrid Feature Selection Approach

Abstract

The prevalent use of social media facilitates the spreading of malicious messages that may inject misleading information to divert normal market operations. However, identifying these threats from noisy social media poses a challenge due to the rarity of confirmed positive threat cases and severe imbalance of ordinary datasets. It is critical to resolve these problems by developing automated techniques that learn from the data and parse out the potential threats. In this paper, we propose a hybrid approach that combines feature selection and resampling techniques for identifying cyber threats from abnormal stock movements. This hybrid feature selection approach (HFS) (1) balances the class distribution and removes the noise from the dataset by using the SMOTE (Synthetic Minority Over-sampling Technique) algorithm and the Tomek Link, (2) identifies the top k best features by using the hybrid feature selection, and (3) further improves model performance by using the subsampling strategy. To test the approach, we collected the social media messages and stock prices of four technology companies: Apple, Microsoft, Intel and Cisco from July 11, 2017 to September 27, 2017. The data consists of 2,244 scenarios, each modeled as a feature vector. We believe that HFS outperforms other state-of-the-art methods for addressing class imbalance problem, and that HFS can effectively support machine learning techniques to detect abnormal stock movements. This research should contribute to the identification of cyber threats from financial social media.