Link prediction and feature relevance in knowledge networks: A machine learning approach

Area of Science:

• Social Sciences
• Physical and Engineering Sciences
• Life Sciences

Background:

• University partnerships are crucial for successful joint research and development.
• The Horizon 2020 programme funds collaborative R&D projects across diverse scientific domains.
• Understanding factors influencing partnership formation is key to fostering innovation.

Purpose of the Study:

• To develop a supervised machine learning model for predicting university partnership formation.
• To analyze the impact of network (endogenous) and non-network (exogenous) features on collaboration prediction.
• To identify key features driving successful joint R&D project formation.

Main Methods:

• Supervised machine learning for link formation prediction.
• Feature importance analysis using super-learner partial effects and elasticities.
• Cross-validation accuracy assessment in two settings: with all features and with exogenous features only.

Main Results:

• Achieved 91% prediction accuracy when including both network and non-network features.
• Prediction accuracy dropped to 67% when using only non-network features.
• Existing program participants (incumbents) showed 24% higher predictive power than newcomers.

Conclusions:

• Network attributes (endogenous features) are most influential in predicting university partnerships.
• The probability of link formation decreases with feature changes, uniformly across attributes and domains.
• Existing collaborations significantly enhance the predictive power of partnership formation models.