Twitter alloy steel disambiguation and user relevance via one-class and two-class news titles classifiers

Abstract

This paper addresses the nontrivial task of Twitter financial disam- biguation (TFD), which is relevant to filter financial domain tweets (e.g., alloy steel or coffee prices) when no unique identifiers (e.g., cashtags) are adopted. To automate TFD, we propose a transfer learning approach that uses freely labeled news titles to train diverse one-class and two-class classification methods. These include different text handling transforms, adaptations of statistical measures and modern machine learning methods, including support vector machines (SVM), deep autoencoders and multilayer perceptrons. As a case study, we analyzed the domain of alloy steel prices, collecting a recent Twitter dataset. Overall, the best results were achieved by a two-class SVM fed with TFD statistical measures and topic model features, obtaining an 80% and 71% discrimination level when tested with 11,081 and 3,000 manually labeled tweets. The best one-class performance (78% and 69% for the same test tweets) was obtained by a term frequency-inverse document frequency classifier (TF-IDFC). These models were further used to gen- erate a Financial User Relevance rank (FUR) score, aiming to filter relevant users. The SVM and TF-IDFC FUR models obtained a predictive user discrimination level of 80% and 75% when tested with a manually labeled test sample of 418 users. These results confirm the proposed joint TFD-FUR approach as a valuable tool for the selection of Twitter texts and users for financial social media analytics (e.g., sentiment analysis, detection of influential users).