MLNE: Multi-Label Network Embedding.

Area of Science:

• Graph mining and machine learning
• Network representation learning
• Data science

Background:

• Network embedding methods learn low-dimensional vector representations of network structures.
• Existing methods struggle with instances possessing multiple labels (multi-labels).
• Multi-label data is common in real-world complex systems.

Purpose of the Study:

• To formulate and address the multi-label network embedding (MLNE) problem.
• To develop a framework that integrates network topology, node content, and multi-label correlations.
• To improve representation learning for networked instances with multiple labels.

Main Methods:

• Proposed a two-layer network embedding framework.
• Constructed a high-level label-label network alongside a low-level node-node network.
• Utilized a unified training objective to optimize both node and label representations in a shared latent space.

Main Results:

• The proposed MLNE framework effectively captures higher-order label correlations.
• Latent label-specific features enhance the low-level node network.
• Experiments show superior performance compared to existing methods on real-world datasets.

Conclusions:

• MLNE provides an effective approach for learning representations of networked multi-label instances.
• Integrating topology, content, and multi-label correlations is key to MLNE.
• The proposed framework advances graph mining capabilities for complex, multi-labeled networks.