Online Course Introduction to Graph Algorithms with Neo4j 4.0 Overview of Graph Algorithms Introduction to Graph Data Science library Graph Algorithms Workflow Environment Setup Graph Management Community Detection Algorithms Centrality Algorithms Similarity Algorithms Recipes Analysis Memory Requirements Estimation Additional Information… Read more →

Additional Information

Monopartite graph

Here be dragons… (warning!)

The algorithms are built for monopartite graphs Monopartite = only one type of node Solution: project a monopartite graph (but this only sometimes makes sense) The algorithms will run … even when they shouldn’t Most similarity algorithms rely on feature vectors that are the same length May not always return meaningful results



Neo4j Cluster

GraphDatabase.driver( “bolt+routing://server?policy=OLAP” )

causal_clustering.server_groups=olap1 causal_clustering.load_balancing.config.server_policies.OLAP=groups(olap1)

Types of graphs

Tie the flavors of graph back to what it means for running algorithms:

Directed and weighted graphs have more data that can be used as inputs for your algorithms to tune their performance.

Graph density is going to impact your performance (but is also not really a type of graph) but isn’t an input.

Cyclic/Acyclic graphs have implementations for some algorithms, but we don’t implement any algorithms that require acyclic graphs.

Other tips

Streaming lots of results with the Neo4j Drivers can be very slow. Lots of results = easier to store those and query them afterwards. Or, wrap a streamed result in apoc.export.csv. Avoid writing to core cluster members. Stream data in a cluster, use snapshot to write results. Avoid running graph algorithms on a transactional graph May get unexpected behavior when you write back to the graph. Similarity algorithms return nonsensical results when vectors are of different lengths.

Common concerns

Why stream vs store results? Algorithm takes really long to run. Betweenness Centrality is slow (all pairs shortest path). Community Detection gives different results each time. Resolved by using a “seeding” parameter in Label Propagation and unionFind Do relationship directions matter? Triangle Count – ignores direction. PageRank – an outgoing relationship means you give credibility to the other node. What about weights? Which algorithms cans I use?


Stay Connected

Sign up to find out more about Neo4j's upcoming events & meetups.