Behavior Based Social Dimensions Extraction for Multi-Label Classification.

Area of Science:

• Network Science
• Machine Learning
• Data Mining

Background:

• Multi-label classification in heterogeneous networks often relies on social dimensions.
• Traditional methods using community detection for social dimensions can fail, leading to poor classification performance.
• Existing approaches struggle when community detection algorithms are ineffective.

Purpose of the Study:

• To propose a novel behavior-based social dimensions extraction method for multi-label classification in heterogeneous networks.
• To improve classification performance by utilizing node behavior features instead of community memberships.
• To address the limitations of traditional community detection-based methods.

Main Methods:

• Developed a behavior-based social dimensions extraction technique.
• Employed Latent Dirichlet Allocation (LDA) to model network generation and extract connection behaviors.
• Utilized extracted connection behaviors as latent social dimensions for classification.

Main Results:

• The proposed method demonstrates satisfactory classification results on various public datasets.
• Achieved superior performance compared to state-of-the-art methods, particularly with smaller numbers of social dimensions.
• Node behavior features proved effective for extracting relevant social dimensions.

Conclusions:

• Behavior-based social dimensions extraction offers a robust alternative to community detection for multi-label classification.
• The LDA-based approach accurately models network generation and extracts meaningful social dimensions.
• This method enhances classification accuracy in heterogeneous networks, especially in challenging scenarios.