Examining Supervised Machine Learning Methods for Integer Link Weight Prediction Using Node Metadata.

Area of Science:

• Network analysis
• Machine learning
• Data science

Background:

• Link weight prediction is crucial for understanding network dynamics.
• Node metadata is a potential feature for improving prediction accuracy.
• Previous studies have explored various features for link prediction.

Purpose of the Study:

• To evaluate the impact of node metadata similarity on link weight prediction accuracy.
• To compare the performance of supervised machine learning methods with and without metadata similarity features.
• To investigate the role of node metadata in link weight prediction through synthetic data analysis.

Main Methods:

• Utilized supervised machine learning methods for link weight prediction.
• Incorporated node metadata as a similarity feature.
• Treated link weights as count variables (number of interactions).
• Analyzed four empirical datasets and synthesized data with varying metadata-dependency.

Main Results:

• No significant improvement in prediction accuracy was observed when incorporating metadata similarity in empirical datasets.
• Random forest outperformed other methods in 99.07% of cases with synthetic data where weights solely depended on metadata.
• Prediction accuracy significantly degraded for all methods when weights were solely based on metadata compared to original weights.

Conclusions:

• Node metadata similarity, as implemented, did not substantially enhance link weight prediction accuracy for the studied empirical datasets.
• The effectiveness of node metadata in link weight prediction may depend on the underlying data generation process and network structure.
• Further research is needed to explore alternative methods for incorporating node metadata and its potential in complex network analysis.