The esports industry in Indonesia is rapidly growing and gaining significant attention on social media, particularly YouTube, where comments reflect public perceptions. This study compares the performance of Naive Bayes and Support Vector Machine (SVM) in classifying sentiments from YouTube comments and explores key themes using Latent Dirichlet Allocation (LDA). Data were collected via the YouTube Data API v3, labeled with TextBlob and manually verified into positive, negative, and neutral categories. After preprocessing and TF-IDF representation, class imbalance was handled with SMOTE, and models were trained and evaluated using accuracy, precision, recall, F1-score, and confusion matrix. Results indicate that Naive Bayes achieved 73.85% accuracy with an F1-score of 0.71, while SVM slightly outperformed with 73.97% accuracy and the same F1-score. SVM showed better consistency in classifying negative and neutral comments, whereas Naive Bayes was more effective for positive ones. LDA revealed dominant discussion topics such as appreciation, enthusiasm, community interaction, criticism, and support for esports development. These findings highlight SVM’s superior overall performance and the value of LDA in uncovering public discourse, providing both academic contribution and practical insights for the esports industry in understanding public sentiment.