Machine Learning Classifiers for Twitter Surveillance of Vaping: Comparative Machine Learning Study.

Area of Science:

• Computational linguistics
• Public health informatics
• Machine learning applications

Background:

• Twitter is a valuable platform for public health surveillance of vaping information and sentiment.
• Machine learning classifiers can identify vaping-relevant tweets and sentiments, supporting a Twitter-based surveillance system.
• Deep learning classifiers require fewer annotated tweets than traditional methods, leveraging large unannotated datasets.

Purpose of the Study:

• To derive and evaluate traditional and deep learning classifiers for identifying vaping-relevant tweets.
• To classify tweets based on commercial nature and provape sentiments.
• To assess the performance of various machine learning models in vaping-related tweet analysis.

Main Methods:

• Collected vaping-related tweets over two months (August-October 2018).
• Manually annotated 4000 tweets for relevance, commercial nature, and sentiment.
• Developed traditional classifiers (logistic regression, random forest, SVM, naive Bayes) and deep learning classifiers (CNN, LSTM, LSTM-CNN, BiLSTM).
• Utilized unannotated tweets for word vector generation to enhance deep learning model performance.

Main Results:

• LSTM-CNN achieved the highest AUC (0.96) for relevance classification.
• Deep learning models, including LSTM-CNN, outperformed traditional classifiers in distinguishing commercial tweets (AUC 0.99).
• BiLSTM showed the best performance for provape sentiment classification (AUC 0.83).
• LSTM-CNN demonstrated superior overall performance across all three classification tasks.

Conclusions:

• Deep learning classifiers, especially LSTM-CNN, offer superior performance for analyzing vaping-related tweets.
• Deep learning models require less preprocessing and fewer annotated data points compared to traditional methods.
• The developed classifiers support the creation of an effective Twitter-based vaping surveillance system.