Cluster features and code necessary to train (1) a classic Probit regression model and (2) a linear neural network to identify extreme growth research clusters in the Map of Science
The full list of 141 cluster features used as input for the classifier are provided in Cluster Features.xlsx.
There are 93 features that relate directly to time series information (articles, citations, references), and the remaining 48 are derived features included for a specific reason someone might think they’re relevant.
- Life Cycle: Inspired by stages of innovation, such as peak years, vitality, growth, etc.
- 16 Features (L0 - L15)
- Academic Importance: Looks at top journals from the past couple of years
- 2 Features (I0 - I1)
- Document Counts: Different document types (review, pre-prints, etc.)
- 4 Features (D0 - D3)
- Network: Explores network features of the clusters, such as average weighted or unweighted degree
- 7 Features (N0 - N6)
- Field: Fraction of articles in the 12 top level classification art from the Map of Science
- 12 Features (F0 - F11)
- Patents: Number of citations from patent families
- 1 Feature (P0)
- Constructs: Other metrics identified from papers that involve more elaborate construction (e.g. entropy)
- 3 Features (C0 - C2)
- Countries and Collaborations: How international or collaborative a cluster is
- 3 Features (CC0 - CC2)
- Time series Information: The share of articles from each year assigned to the cluster
- 16 Features (S0 - S15) for growth share info
- 15 Features (dS0 - dS14) for change in growth share info
A full list of funding-critical clusters are provided in the document funding_critical_cluster_metadata.csv