Max weighted matching

[jpweytjens]

This pull request implements max weighted bipartite graph matching as a method for the OneToOne class. It uses networkx for the actual matching algorithm. Multiple helper functions are defined to transform the MultiIndex output from recordlinkage to a graph object and back.

[jpweytjens]

You can easily visualize (small) graphs with the following code

def draw_weighted_bipartite_graph(graph):
    """Draw a weighted bipartite graph. 
    No garantuees are made about the order of the nodes."""

    labels = nx.get_edge_attributes(graph, "weight")
    nodes = [n for n, d in graph.nodes(data=True) if d["bipartite"] == 0]
    pos = nx.drawing.layout.bipartite_layout(graph, nodes)

    nx.draw_networkx_edge_labels(graph, pos, labels)
    nx.draw_networkx(graph, pos)

though I'm not sure recordlinkage should contain such a function?

[codecov]

Codecov Report

Merging #107 into master will increase coverage by 0.54%.
The diff coverage is 95.71%.

@@            Coverage Diff             @@
##           master     #107      +/-   ##
==========================================
+ Coverage    77.2%   77.75%   +0.54%     
==========================================
  Files          33       33              
  Lines        2338     2400      +62     
  Branches      376      391      +15     
==========================================
+ Hits         1805     1866      +61     
+ Misses        408      407       -1     
- Partials      125      127       +2

Impacted Files	Coverage Δ
recordlinkage/network.py	91.96% <95.71%> (+9.27%)	⬆️
recordlinkage/base.py	80.31% <0%> (-0.17%)	⬇️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 87a5f4a...470985f. Read the comment docs.

[jpweytjens]

If I find the time, I'm interesting in adding the symmetric best matching and stable matching methods from Franke et. al as well. These methods seem to produce comparable results, for a much lower computational cost.

[J535D165]

Many thanks for this well written Pull Request. A few change requests and remarks below.

[J535D165]

Thanks @jpweytjens

You can easily visualize (small) graphs with the following code

def draw_weighted_bipartite_graph(graph):
    """Draw a weighted bipartite graph. 
    No garantuees are made about the order of the nodes."""

    labels = nx.get_edge_attributes(graph, "weight")
    nodes = [n for n, d in graph.nodes(data=True) if d["bipartite"] == 0]
    pos = nx.drawing.layout.bipartite_layout(graph, nodes)

    nx.draw_networkx_edge_labels(graph, pos, labels)
    nx.draw_networkx(graph, pos)

though I'm not sure recordlinkage should contain such a function?

That would be nice to have in the documentation or on the examples page. If it turns out to be a succes, we add it to the toolkit.

Documentation: https://recordlinkage.readthedocs.io/en/latest/ref-classifiers.html?highlight=onetoone#network

[jpweytjens]

I've added a new commit that should address all the requested changes.

• A new function add_weights is introduced to add weights to candidate matches with 3 different methods. This function returns a Series instead of DataFrame.
• Series are now used throughout the max_weighted matching function, instead of DataFrames.
• networkx has been removed from tox and added to setup.py
• Documentation has been added for the max_weighted method and an example is given for the new add_weights function.

I'm curious to hear your thoughts on this new version. If you're happy with them, I would like to add a test for the max_weighted method.

[jpweytjens]

Can you review the changes @J535D165 ?