Machine Learning Models Detect Deaths from Social Media Posts
A dissertation on arXiv (2608.05183) explores the automatic detection of deaths from social networking site posts by analyzing linguistic differences between pre-mortem and post-mortem content. The study developed a new dataset using Wikidata and Twitter, and trained machine learning classifiers—both traditional (RF, KNN, LR, SVM) and deep learning (BiLSTM, CNN, BERT)—on features extracted via TF-IDF and pre-trained embeddings (GloVe, Word2Vec, FastText). Results showed RF outperformed other traditional models, BiLSTM outperformed CNN, TF-IDF was superior for traditional models, and Word2Vec excelled for deep learning models. The findings highlight the potential of ML in detecting user deaths from social media, with implications for digital legacy and online memorialization.
Key facts
- Dissertation on arXiv:2608.05183
- Analyzed linguistic differences between pre-mortem and post-mortem social media content
- Developed new dataset using Wikidata and Twitter
- Tested traditional ML models: RF, KNN, LR, SVM
- Tested deep learning models: BiLSTM, CNN, BERT
- Used TF-IDF and pre-trained embeddings: GloVe, Word2Vec, FastText
- RF outperformed other traditional models
- BiLSTM outperformed CNN
- TF-IDF consistently outperformed pre-trained embeddings for traditional models
- Word2Vec outperformed GloVe and FastText for deep learning models
Entities
Institutions
- arXiv
- Wikidata