Performance Comparison of One-Hot Encoding vs Response Coding in Context of Social Media Text Sentiment Analysis

Most of the content on various social media platforms has enormous textual data. Before being used in machine learning models, this textual data must be transformed into numerical formats using feature extraction and feature engineering techniques to improve classification accuracy. Numerous methods, including Label Encoding, Response Coding, One-hot Encoding, and Binary Encoding, are available for converting textual input into numerical values, and these methods are effective in handling massive amounts of textual data. This study investigates how these methods can be applied to social media reviews and comments to establish the most effective encoding technique for classification machine learning models that recognize sentiment polarity. We selected a social media dataset with 44,955 tweets for our work and divided it into train, test, and cross-validation. Experimental results show that One-hot encoding performed the best with Logistic regression, whereas Response coding was less efficient.