Sequential stacking link prediction algorithms for temporal networks.

Area of Science:

• Network Science
• Data Mining
• Machine Learning

Background:

• Link prediction algorithms are crucial for network analysis, aiding data collection and inferring missing connections.
• Dynamic networks, where links evolve over time, present challenges for traditional link prediction methods.
• Optimally leveraging temporal information in link prediction remains an open research question.

Purpose of the Study:

• To evaluate the effectiveness of temporal topological features versus static network features for temporal link prediction.
• To develop and validate a novel sequential stacking method for improved temporal link prediction.
• To assess the performance of the proposed method across various network types and data completeness scenarios.

Main Methods:

• A sequential stacking approach utilizing 41 static network features was employed, minimizing manual feature engineering.
• The method was tested on temporal stochastic block models and 19 real-world temporal networks.
• Ensemble learning by combining the proposed method with other predictors was used to enhance performance.

Main Results:

• Sequentially stacked static network features demonstrated higher accuracy than many temporal topological features in link prediction.
• The proposed method achieved near-oracle-level performance on temporal stochastic block models, especially in ensemble configurations.
• Performance improvements were consistently observed across diverse real-world temporal networks.

Conclusions:

• Sequential stacking of static network features offers a computationally efficient and accurate alternative for temporal link prediction.
• The developed method effectively handles both partially and fully unobserved network layers.
• Ensemble learning by stacking multiple predictive methods significantly boosts performance in real-world temporal network analysis.