Update: This paper is now published in the Journal of Physics: Complexity – https://iopscience.iop.org/article/10.1088/2632-072X/aba83d.
Update: This paper is now on ArXiv at https://arxiv.org/abs/2001.01126. I just returned from the Social Networks & Health Workshop at Duke University, where I presented my research on joining graph and language embeddings to predict individuals at risk of suicide! The presentation went really well, and it was great to see the work that other Fellows had done since last year. I highly recommend the Workshop to anyone wanting to learn more about advanced social network analyses and how they can be applied to study different health outcomes. If you want to know more, please get in touch with me! The project I discussed is work I am doing for my PhD qualifying exam. It demonstrates my competencies in key areas of computational social science, social network analysis, and social informatics. I show how both graph and document embeddings generate good predictive results of whether a Reddit user will post in SuicideWatch; however, both models return quite a few false-positives and false-negatives. When used in conjunction, on the other hand, the model returns few false-positives and false-negatives and has a F1-score above 0.90. Pearson correlations between the graph and two document embeddings (from doc2vec distributed bag of words and distributed memory models) show that each model is tracking along different information even though they have solid overlap. The graph embeddings, for example, only have approximately 0.2-0.25 correlation with the two document embeddings, and the two document embeddings have a correlation of 0.6. Very excited to talk about these results more when my paper is done and up on arXiv!
0 Comments
|
|