Author :
Abstract
Cryptocurrencies designed to be used on the internet have attracted the attention of many people since the day they were released. As the number of people turning to cryptocurrency trading increased, the number of texts on the subject also increased and the need to analyze the texts, which turned into a large data set, arose. There are analysis studies in the literature using tweets taken from the Twitter platform on different topics. This study aims to perform sentiment analysis on cryptocurrency-tagged tweets shared on the Twitter platform, in line with the rising popularity of cryptocurrencies and the interest of internet users in this field. Since Twitter is a popular social media platform that offers a wide data set, tweets containing users' emotional expressions about cryptocurrencies constitute the main data source of this study, and a data set specific to the study was created by automatically tagging tweets with LSTM. Feature detection was made with LDA and NMF algorithms, which are topic modeling methods; Models were created with NB, LR and SVM algorithms and results were obtained for classification accuracy and f1-score metrics. When the results were compared, the highest success was achieved when the LDA method and SVM classifier were used, and 87.89% accuracy and 87.85% F1-criterion values were reached in the model created by setting the subject feature size as 500 and the number of components as 40. The best result obtained with the NMF method was 87.43% accuracy and 87.34% F1-criterion with the model created with the SVM classifier when the subject feature size and the number of components were 500 and 50, respectively. The second-best performance was obtained with the LR classifier for both methods; With the LDA method, 87.13% accuracy was recorded when the subject feature size and the number of components were selected as 300 and 40, and 87.13% F1-criterion values were recorded when 500 and 30 were selected, respectively, while with the NMF method, 87.01% accuracy and 86.90% F1-criterion values were reached when the subject feature size was 500 and the number of components was 50. The lowest performance was obtained with the NB classifier for both methods. The NB classifier showed its best performance with 73.00% accuracy and 73.01% F1-measure values with LDA when the subject feature size and the number of components were selected as 200 and 50, respectively, and 70.00% accuracy and 70.09% F1-measure values with NMF when 500 and 40 were selected. According to the results, both LDA and NMF methods generally showed successful performance. The highest and most consistent performances were obtained with the SVM classifier in both methods, and it was concluded that the SVM algorithm was the most successful classifier compared to the other two classifiers used. In addition, the findings obtained from the study show that emotional expressions related to cryptocurrencies on Twitter can be analyzed effectively.
Keywords
Abstract