Week 1
Literature reviews, overview of the project.
Last updated
Was this helpful?
Literature reviews, overview of the project.
Last updated
Was this helpful?
the idea is to use the dynamic measures to see the graph
for HD features or combined features, using a connected scatter plot seems interesting
narrow down the design, and focus more on the problem itself rather than users
time: snapshot → period → lifetime
component: local → subgraph → global
local
subgraph
global
snapshot
single value
single value
single value
period
aggr, series, derived
aggr, series, derived
aggr, series, derived
lifetime
aggr, series, derived
aggr, series, derived
aggr, series, derived
"aggregated value" refers to the calculated metric value from an aggregated graph. "series value" means a value sequences for each snapshot during the period. And "derived value" is the derived value from series value, such as extremes, average value, deviance, trend, stability, fastest increase, etc.
Global
Static
count, ratio
number of nodes, active nodes, links, node pairs, connected components, motifs (triads, cliques...), clusters, activation, redundancy
clustering coefficient
extreme
diameter,
average
characteristic path length ( shortest path length), efficiency (1/shortest path)
others
modularity
Dynamic
global volatility
Local
Static
centrality: identify the most important vertex or link in a graph
degree, strength (sum of weight)
eccentricity (max d_ij)
closeness (average shortest path length)
betweenness
others
pagerank, redundancy
Dynamic
volatility: sum presence variance of links to a node
activation: count of new connections
High-Dimensional Features (learned from data)
Deepwalk, Graph2Vec, SDNE, matrix factorization, persistent diagram, adjacent matrix