SCALABLE TWITTER (X) SENTIMENT ANALYSIS USING TF-IDF AND CHI-SQUARE FEATURE OPTIMIZATION WITH WEIGHTED SUPPORT VECTOR MACHINES

Abstract

Abstract - In this paper, Weighted Support Vector Machine (WSVM) architecture is presented as a model that is adapted to issues of large-scale sentiment classification of Twitter(X) data. The paper solves the most common issues that are associated with social media text: lexical noise, sparse feature representations, and very high-dimensional feature spaces, using the publicly available Sentiment140 corpus, which contains 1.6 million annotated tweets. It suggests a hybrid feature engineering pipeline, which combines optimised Term Frequency Inverse Document Frequency (TF-IDF) weighting of both unigram and bigram n-grams, and then chi-square statistical feature selection. The classification step uses a Radial Basis Function (RBF) kernel SVM trained with cross-validated grid search, and class-imbalance mitigated by instance-level sample weighting. The final proposed WSVM offers a 88.5, 87.9, 87.2 and 87.6 percent accuracy, precision, recall, and F1-score respectively on the held-out test partition, respectively, which are statistically significant improvement over baseline SVM, Naive Bayes, Logistic Regression, and random forests classifiers. Findings show that principled feature engineering and class-aware training methods provide competitive performance without the computational cost of deep transformer models.