Research Ranking
Skills & Tools
Python
Doc2Vec
Web-scraping
XGBoost
Random Forests
Neural Networks
Variational Autoencoder
Isomap
Predicting Important Papers Early
Especially in times of crisis, with a high volume of publications and a premium on speed, knowing where to look is important. Using previous publication and citation data, we are seeking to flag impactful research about COVID-19 as it emerges. We hope to focus literature reviews for medical professionals so they can focus on their patients instead of piles of paper.
Our inquiry shows that it is possible to learn about citation count from these data. However, it is also clear that the majority of the signal comes from non-textual features, the date and journal of publication were much more predictive. This result is not conclusive though. Since our data are a subset of the total literature available on coronaviruses, our features designed to capture the relationship to other work may be impacted. This problem may also lead to inflated importance of underrepresented journals. If a journal is only represented a handful of times with highly cited papers, this may inflate the journal’s importance if the papers unaccounted for have low citation counts. For this reason, it would be interesting to include analysis that has already been conducted on the entire body of works for each author and journal, like the h-index.
Taking this into account, it is likely that our text analysis and embedding techniques are not well suited to this task. Typically this type of work deals with disparate fields of study, while this inquiry looks at a set of academic papers focused on closely related subjects. This inherit similarity makes it more difficult for topic modeling and document vectorization techniques to separate the papers.
While our current models performed moderately well, we believe using tree based predictive modeling in combination with embedded texts has more promise on a more diverse set of papers. Interesting further research would include testing these methods on a more complete body of coronavirus literature and on papers from a wider range of academic fields.