splink-graph 0.8.2

a small set of graph functions to be used from pySpark on top of networkx and graphframes

splink_graph: Graph metrics for data linkage at scale

---

splink_graph is a graph utility library for use in Apache Spark.

It computes graph metrics on the outputs of data linking which are useful for:

• Quality assurance of linkage results and identifying false positive links
• Computing quality metrics associated with groups (clusters) of linked records
• Automatically identifying possible false positive links in clusters

It works with graph data structures such as the ones created from the outputs of data linking - for instance the candidate pair results produced by

Calculations are performed per cluster/connected component/subgraph in a parallel manner thanks to the underlying help from pyArrow.

---

TL&DR :

Graph Database OLAP solutions are a few and far between. If you have spark data in a format that can be represented as a network/graph then with this package:

• Graph-theoretic metrics can be obtained efficiently using an already existing spark infrastucture without the need for a graph OLAP solution
• The results can be used as is for finding the needle (of interesting subgraphs) in the haystack (whole set of subgraphs)
• Or one can augment the available graph-compatible data as part of preprocessing step before the data-ingestion phase in an OLTP graph database (such as AWS Neptune etc)
• Another use is to provide support for feature engineering from the subgraphs/clusters for supervised and unsupervised ML downstream uses.

How to Install :

Note that splink_graph 0.8.2 is suitable for Spark 3.x only

The easiest way to install splink_graph 0.8.2 is to type

pip install splink_graph==0.8.2 in your terminal

If you are interested in running splink_graph in a Spark 2.4.x environment then type

pip install splink_graph==0.5.0 . Codewise all 0.5.0 code is located at the splink_graph_0_5_0 branch

For dependencies and other important technical info so you can run these functions without an issue please consult INSTALL.md on this repo, as for each Spark version there are specific prerequisite actions you might need to take in order to not face issues.

Functionality offered :

For a primer on the terminology used please look at TERMINOLOGY.md file in this repo

Cluster metrics

Cluster metrics usually have as an input a spark edgelist dataframe that also includes the component_id (cluster_id) where the edge is in. The output is a row of one or more metrics per cluster

Cluster metrics currently offered:

• diameter (largest shortest distance between nodes in a cluster)
• transitivity (or Global Clustering Coefficient in the related literature)
• cluster triangle clustering coeff (or Local Clustering Coefficient in the related literature)
• cluster square clustering coeff (useful for bipartite networks)
• cluster node connectivity
• cluster edge connectivity
• cluster efficiency
• cluster modularity
• cluster assortativity
• cluster avg edge betweenness
• cluster weisfeiler lehman graphhash (in order to quickly test for graph isomorphisms)

Cluster metrics are really helpful at finding the needles (of for example clusters with possible linking errors) in the haystack (whole set of clusters after the data linking process).

---

Node metrics

Node metrics have as an input a spark edgelist dataframe that also includes the component_id (cluster_id) where the edge belongs. The output is a row of one or more metrics per node

Node metrics curretnly offered:

• Eigenvector Centrality
• Harmonic centrality

---

Edge metrics

Edge metrics have as an input a spark edgelist dataframe that also includes the component_id (cluster_id) where the edge belongs. The output is a row of one or more metrics per edge

Edge metrics curretnly offered:

• Edge Betweeness
• Bridge Edges

---

Functionality coming soon

• release for MVP to be used on AWS glue and demos
• cluster modularity based on partitions created by edge-betweenness
• cluster number of bridges metric added
• cluster assortativity added
• cluster modularity based on partitions created by label propagation
• shallow embeddings of subgraphs/clusters (WIP)
• Add a connected components function (from the graphframes library)
• Add a connected components function for smaller graphs (from the networkx library) so its easier to get started.

For upcoming functionality further down the line please consult the TODO.md file

Contributing

Feel free to contribute by

• Starting an issue.

• Forking the repository to suggest a change, and/or

• Want a new metric implemented? Open an issue and ask. Probably it can be.