EvolveGCN: Evolving Graph Convolutional Networks for Dynamic Graphs
Graph Deep Learning
We live in a networked world of complex relationships. Graph structured data captures these relationships (edges) between entities (nodes), as well as their associated properties. Think about your social network, a supply chain network, or even a Wikipedia knowledge graph. Each contains various patterns, interdependencies, and insights that can be revealed with proper context. Deep learning on graphs has lagged other segments of AI because the combinatorial complexity and nonlinearity of graphs requires long training times. Recently IBM Research and others have made big steps forward on scalability, triggering an exciting acceleration in the field. Graph learning is powerful for industry applications. A financial institution can analyze a transaction network to detect sophisticated money-laundering activity. A biotech firm can analyze protein- expression networks to understand the molecular basis of a disease. A manufacturing giant can analyze its massive supply chain to optimize inventory management and monitor risk. In each case, the fundamental algorithmic work must be done first – and we’re on it.
Anti-Money Laundering in Bitcoin: Experimenting with Graph Convolutional Networks for Financial Forensics