Dynamic Co-Embedding Model for Temporal Attributed Networks

Area of Science:

• Computer Science
• Data Science
• Network Science

Background:

• Temporal attributed networks require methods to capture dynamic node-attribute relationships.
• Existing temporal network embedding methods fail to represent dynamic node-attribute affinities.
• Static co-embedding methods are not directly applicable for dynamic temporal attributed networks.

Purpose of the Study:

• To develop a novel model for learning dynamic low-dimensional representations of temporal attributed networks.
• To address limitations in existing methods regarding dynamic node-attribute affinities.
• To enable tracking of network evolution and preservation of node-attribute relationships over time.

Main Methods:

• Proposed the dynamic co-embedding model for temporal attributed networks (DCTANs).
• Utilized a dynamic stochastic state-space framework to model belief states and transitions.
• Learned embeddings for both nodes and attributes at each time step.

Main Results:

• DCTANs effectively capture the dynamics of temporal attributed networks.
• The model preserves affinities between nodes and attributes over time.
• Achieved substantial performance gains in static and dynamic graph mining tasks compared to state-of-the-art models.

Conclusions:

• DCTANs provide a robust framework for embedding temporal attributed networks.
• The model successfully learns dynamic node and attribute representations.
• Demonstrated superior performance in real-world graph mining applications.