Chaining algorithms predict the emergence of scientific ideas

To reproduce the figures and analysis in this paper:

Collecting data given a list of scientists

1. Run query_semanticscholar.py on the domain {CS, Chemistry, Economics, Medicine, Physics}, or your own list of scientists
2. Run filter_by_year using the birth/death dates file, output to abstracts_filtered_year
3. Clean the abstracts with get_vectors.py, output to abstracts-cleaned directory
4. Encode the abstracts with SBERT via sbert.py, output to sbert-abstracts
5. Order/convert dates to timestamps via emergence_order.py, output to abstracts-ordered directory
6. Run models on the abstracts-ordered directory

Running models

1. Hyperparameters must be tuned for models on each scientist: run opt_hyperparam_exemplar.py for each model/field combination, output individual param values to `/individual-s-vals/

2. Run comparison between models: src/models/predict.py --type <nobel/turing> --field <field> --measure ll -i

3. Run shuffle tests between models: src/models/predict.py --type <nobel/turing> --field <field> --measure ll -i -s --sy

4. Run authorship analysis: src/models/predict_k_author_papers.py --type <nobel/turing> --field <field> -k <max authors, or -1 for first author>

Creating figures

Most figures are generated through functions in rain_plots.py, based on simulation outputs generated through the "Running models" section. The stacked authorship charts can be generated through stacked_bar_authorship.py. tSNE visualizations can be generated through make_tsne_figure.py.