Inferring Users' Social Roles with a Multi-Level Graph Neural Network Model.

Area of Science:

• Social Network Analysis
• Machine Learning
• Data Science

Background:

• Social networks feature diverse user roles (e.g., celebrities, officials, organizations) and professional statuses (e.g., managers, workers).
• Prior research often relies on categorical, textual, and topological data for role prediction, which may not capture comprehensive user characteristics.
• Existing methods have limitations in fully reflecting the multifaceted nature of social statuses and roles within network structures.

Purpose of the Study:

• To investigate how social network structures inherently reflect users' social statuses and roles.
• To develop an advanced social network representation learning algorithm for inferring user roles.
• To improve the accuracy of predicting social statuses and roles by incorporating dynamic behavior features.

Main Methods:

• Analyzed a preprocessing mechanism to extract dynamic behavior features from social network datasets, using the Enron email dataset as a case study.
• Designed a novel social network representation learning algorithm utilizing an attention and gate mechanism on user neighbors.
• Employed four publicly available datasets for extensive experimental validation.

Main Results:

• The proposed algorithm demonstrated superior performance in inferring users' social statuses and roles.
• Achieved an average accuracy improvement of 2% over GraphSAGE-Mean, a leading inductive representation learning method.
• Validated the effectiveness of incorporating dynamic behavior features and attention mechanisms in social network analysis.

Conclusions:

• Social network structures provide valuable insights into users' social statuses and roles.
• The novel representation learning algorithm effectively infers user roles by analyzing network structure and dynamic behaviors.
• This approach offers a significant advancement in social network analysis for understanding user characteristics and improving prediction accuracy.