EVALUATION OF THE NEURAL NETWORK MODEL'S EFFICIENCY IN ANALYZING THE USERS' EMOTIONAL MOOD IN SOCIAL NETWORKS

Abstract

This study investigates the effectiveness of various neural network models for sentiment analysis, specifically for detecting the emotional states of users based on textual data from social media platforms, focusing on Twitter posts related to COVID-19. The research evaluates the performance of six models: Convolutional Neural Network (CNN), CNN with the Adam optimizer, Random Forest, Extra-tree, XGBoost, and BERT. The models were assessed using key evaluation metrics such as accuracy, precision, recall, F1-score, and confusion matrices to measure their ability to classify positive and negative emotions in user-generated text.

The experimental results demonstrated that the BERT model achieved the highest performance, with % overall accuracy of 86%. This result highlights BERT’s superior ability to accurately classify emotions with minimal misclassification, making it an optimal choice for sentiment analysis in social media data. BERT’s performance can be attributed to its deep contextual understanding of the text, which is essential for accurately interpreting emotional tones in complex language.

The CNN and CNN with the Adam optimizer models achieved a general accuracy of 83%, slightly lower than BERT but still outperforming the traditional machine learning models such as Random Forest and XGBoost. Notably, including the Adam optimizer improved the F1-score for the positive class, indicating better performance in detecting positive sentiments. Despite this improvement, these models still lag behind BERT regarding overall accuracy and efficiency.

In contrast, Random Forest, Extra-tree, and XGBoost models showed significantly lower results. Specifically, Random Forest and XGBoost exhibited poor accuracy in detecting positive emotions, with 57% and 54% precision scores, respectively. It suggests limited effectiveness in distinguishing between positive and negative emotions in the analyzed texts. Extra-tree, while outperforming Random Forest and XGBoost, still failed to reach the level of accuracy demonstrated by CNN and BERT, with a general accuracy of 68%.

The findings of this study confirm that deep learning models, particularly BERT, are highly effective for sentiment analysis tasks, especially when working with complex social media data. These models can capture intricate patterns and contextual nuances that traditional machine learning models struggle to identify, making them the preferred choice for tasks related to emotion detection in text. The research underscores the need to select appropriate algorithms based on the specific characteristics of the dataset and task requirements.