Assessment of Iterative Semi-Supervised Feature Selection Learning for Sentiment Analyses: Digital Currency Markets

Currently, the cryptocurrencies have been highly attracted by investors. Due to the lack of a central money authority, the stock markets are experiencing large price fluctuations based on speculation. In this study, our goal is to try to determine the effects of social media topics in digital money markets by using sentiment analysis methods on natural data. First, related comments on Twitter about digital currencies were collected by using a web crawler. The comments investigated with feature selection metrics, semi-supervised feature selection metrics, Random Forest, Support Vector Machine (SVM) methods. The Random Forest was observed to perform the most successful and fastest sentiment classification process using a combination of semi-supervised feature selection metrics by iterative actions. Iterative Semi-Supervised Feature Selection (ISSFS) method proposed and evaluated in this paper. As a result of the experiments, the optimum number of words for the English language was calculated in order to perform sentiment analysis of digital currency markets. Classifiers and methods used were shared as results.

I. Introduction

Studies on sentiment analysis (SA) are now frequently performed today. Thanks to the experience gained in academic and private sector studies, more successful analyses can now be applied on the text mining area. Machine learning algorithms, which produce successful results in most areas, are also frequently used in text mining. Due to the large number of data to be analyzed in text mining, these algorithms are not considered to be successful when used alone due to data pollution. Common methods to overcome this problem; it is carried out by taking a random number of sample words, or using more advanced methods such as S3VM [1], Chi-Square or Information Gain. In this way, the detection of words containing the weight of sentiment from the user messages examined can be carried out more objectively. When the studies about sentiment analysis are examined, the analysis is carried out through a single training set, often dependent on the same words without updating these methods. Since this is the case, considering the intertwining of languages in the globalized world, adding new words to the languages we use every day can adversely affect the success of such systems.