Machine learning prediction of academic collaboration networks

Area of Science:

• Covers European university collaborations across Social Sciences and Humanities (SSH), Physical and Engineering Sciences (PE), Life Sciences (LS), and multidisciplinary research.

Background:

• Existing research has not simultaneously compared network and non-network attributes in collaboration network formation.
• Understanding factors influencing scientific collaborations is crucial for knowledge production and transmission.

Purpose of the Study:

• To investigate the distinct roles of network and non-network attributes in the formation of European university collaborations.
• To compare the predictive power of network versus non-network attributes on collaboration formation across different scientific domains.

Main Methods:

• Employed four machine learning algorithms: LASSO, Neural Network, Gradient Boosting, and Random Forest.
• Analyzed European university collaboration data from 2011 to 2016 across four European Research Council (ERC) domains.
• Evaluated link formation accuracy and feature importance scores.

Main Results:

• The best models achieved over 80% accuracy in predicting link formation.
• Public funding emerged as a significant non-network attribute in PE and LS domains.
• Network attributes were found to be more influential than non-network attributes, substantially increasing prediction accuracy.

Conclusions:

• Network attributes play a more critical role in shaping European university collaborations than non-network attributes.
• Different scientific domains exhibit unique collaboration patterns, indicated by varying feature importance scores.
• Findings highlight distinct modes of knowledge production and transmission within diverse scientific communities.