Public Sentiment Analysis Towards Bitcoin and Ethereum On Platform X Using Graph Neural Network (GNN)

Abstract

The rise in Bitcoin and Ethereum is well-known. Research shows that public sentiment greatly affects their price changes. Thus, public sentiment analysis is a key factor in making investment decisions. This study analyzes public sentiment towards Bitcoin and Ethereum on platform X using Graph Neural Networks (GNN), specifically the GCN and AGN-TSA models. GCN is utilized for its ability to capture syntactic relationships between words, while AGN-TSA integrates textual content with user-level social interactions through an attention mechanism. The dataset is collected from$X$using the keywords Bitcoin, BTC, Ethereum, and ETH, followed by preprocessing and labeling based on FinBERT and Vader as a benchmark labeling technique to construct the training and test sets. Evaluation employs a confusion matrix to compare model performance. The results show that GCN without addressing class imbalance achieves 83.89 % accuracy, whereas AGN-TSA achieves 86.21%. However, confusion matrix analysis revealed severe bias toward the majority class, so we needed to use extreme class weighting (ratio 17:1:10), which improved minority-class recall. However, it caused training instability and reduced accuracy: GCN dropped to 80.58 %, and AGN-TSA dropped to 86.15% (a decrease of$0.06 \%)$. Despite the decrease, AGN-TSA still achieves the best accuracy compared to GCN, this result reinforcing our initial hypothesis that attention-based graph modeling which leverages social ties yields superior performance for sentiment classification in crypto related discourse. Furthermore, the accuracy of the data labeling technique and the imbalance in the label distribution also affect the final accuracy results.