Sentiment analysis and prediction on twitter data using machine learning methods

In this study, an application capable of performing sentiment analysis on text-based data and making predictions using machine learning algorithms was developed. The application was implemented using the Python programming language. A randomly but homogeneously selected subset of 10,000 data points was used from a large dataset of English tweets obtained from the Kaggle platform. Of these data, 70% were allocated for training and 30% for testing, including only examples labeled with positive and negative sentiments. During the data preprocessing stage, URLs, special characters, emojis, and unnecessary words were removed, and words were reduced to their root forms. The cleaned data were converted into numerical vectors using the TF-IDF method and structured with n-gram techniques. The algorithms SVM, Random Forest, and Naive Bayes were applied individually, and hybrid models were also developed using methods such as PCA, Kmeans, and Chi-Squared. By utilizing the GridSearchCV method, the optimal parameters were determined, resulting in the creation of a total of nine different models. When the model performances were compared, it was observed that the hybrid models achieved higher accuracy rates compared to the individual models. The results demonstrated the effectiveness of machine learning algorithms in text mining and sentiment analysis. This study aims to contribute to future research by presenting various modeling strategies.