Twitter’s informal language and changing content present a challenge to established sentiment analysis techniques. This paper presents a unique graph-based method for improving sentiment analysis accuracy using Twitter data. We create graphs from tweets, with nodes representing words or phrases and edges showing relationships. Our approach combines a variety of techniques, including Heterogeneous Graph Neural Networks (Hete-GNNs), Gaussian Naive Bayes using Node2vec, a combined BERT-LSTM-CNN model(Transformer model), and hybrid approaches such as KNN+XGBoost and CNN+LSTM. These approaches produce rich vector representations that are fed into Artificial Neural Networks for sentiment prediction. Our approach outperforms established methods on different datasets, providing a reliable solution for real-time sentiment analysis in the ever-changing social media scene.