- Important transitions
- Conclusions
- Working Parts without thinking about external application
Working Parts:
- networks don’t have spatial locations
- graph neural networks
- node level predictions
- by predicting the likelihood of a node, can predict structures of proteins, drugs side effects, products a person might like (based on knowing attributes of other choices)
- subgraph predictions
- can predict how long a commute might take by adding together predicted time travel from smaller sub-components of graph (road segments)
- graph prediction
- physics simulations - can predict positions and velocities of particles in the future (based on knowing particles and interactions between particles)
- Types of Graphs
- undirected graphs have symmetrical links
- directed graphs have nodes with in degrees and out-degrees
- bipartite graph: two sets or types of nodes, and nodes of one type don’t interact with one another (ie. customers and products, customers don’t have interactions, nor do products)
- multigraph: multiple edges can connect nodes. Can also represent this by edges having weights
- nodes, edges and graphs can all have properties
- connectivity
- strongly: every nodes has a path from one node to another (and vice-versa)
- weakly: connected if we ignore direction of the edge
- SCC: Strongly Connected Components: parts of a weakly connected graph that are strongly connected
- Adjacency Matrix: a way to represent the connections of nodes in a graph to other nodes in a graph
- symmetrical if the graph is undirected
Node Centrality