Neighbourhood Blocking for Record Linkage

Summary

The code in this repository is intended for use with the recordlinkage package. It contains an implementation of the Neighbourhood Blocking Indexing method.
This method combines some features of Standard Blocking, and Sorted Neighbourhood Indexing. In addition, it also allows for meaningful treatment of missing values, the simultaneous use of multiple sorting orders for the blocks (ie: sorting by each of the blocking keys individually) and allowance for a limited number of blocking key mismatches.

IPython Notebooks in this repo perform numerical experiments in which Neighbourhood Blocking is compared to Standard Blocking and Sorted Neighbourhood Indexing. These take some time to run. Their results are that under the conditions tested:

• Neighbourhood has similar scalability properties to Sorted Neighbourhood Indexing, Standard Blocking and Full Indexing (ie: runtime is approximately linear with respect to the size of the index produced)
• Compared to the other methods, Neighbourhood Blocking can produce superior index quality at the expense of increased runtime.

Contents

File	Description
neighbourhood_blocking.pdf	Paper describing Neighbourhood Blocking
neighbourhood_blocking.py	Implementation of Neighbourhood Blocking (requires recordlinkage)
experiment_helpers.py	Helpers used by the test scripts
index_quality.ipynb	Notebook to run index quality test and display results
scalability.py	Script to run scalability test
scalability_results.ipynb	Notebook for viewing results of scalability test
standard_datasets.py	Script to run index quality tests on benchmark datasets
standard_datasets_results.ipynb	Notebook for viewing results of index quality tests on benchmark datasets.

Instructions

To use Neighbourhood Blocking

Use the NeighbourhoodBlockIndex class in neighbourhood_blocking.py. It has the same API as indexers in recordlinkage.

To run the index quality test

In Jupyter, open index_quality.ipynb and run all cells

To run the scalability test

Disable memory paging. To do this on Linux, use:

sudo watch --interval 500 swapoff -a

Then, run scalability.py repeatedly until the file timings.pickle is created (until all tests are done, partial results are saved to and recovered from _timings.pickle). To do this on Linux, navigate the same directory as scalability.py and then use:

while [ ! -f timings.pickle ]; do python scalability.py ; done

When this is complete, results can be viewed by opening scalability_results.ipynb in Jypyter and running all cells.

To install dependencies and download benchmark datasets

Navigate to the root folder and use:

make data

To run index quality tests on benchmark datasets

First download the benchmark datasets as described above. Then disable memory paging as described above and use:

while [ ! -f timings.pickle ]; do python standard_datasets.py ; done

Results can then be viewed using standard_datasets_results.ipynb.