Confusion with the number of points in a node using plot_static_mapper_graph and graph= pipe.fit_transform

[jnukpezah]

Confusion with the number of points in a node using plot_static_mapper_graph and graph= pipe.fit_transform

[ulupo]

Hi! Thanks for the details. Would you be able to produce a minimal code example which illustrates the issue?

[jnukpezah]

Hi!, Yes I can. The code example is below. Thanks.

Define filter function

filter_func = umap.UMAP(n_neighbors=5)

Define cover

cover = CubicalCover(kind='balanced', n_intervals=10, overlap_frac=0.2)
#cover = CubicalCover(n_intervals=10, overlap_frac=0.2)

Choose clustering algorithm

clusterer = DBSCAN(eps=10)

Initialise pipeline

pipe = make_mapper_pipeline(
filter_func=filter_func,
cover=cover,
clusterer=clusterer,
verbose=True,
n_jobs=-1,
)

Plot Mapper graph where df_refined is a Pandas dataframe

fig = plot_static_mapper_graph(pipe, df_refined, color_data = df_refined)
fig.show(config={'scrollZoom': True})

#Node_elements from attributes.
graph = pipe.fit_transform(df_refined)
node_elements = graph.vs["node_elements"]

[jnukpezah]

When I hover over the nodes in the plot_static_mapper graph and match the same node in the node_elements, the node sizes are different. Thanks

[ulupo]

Thanks @jnukpezah! I can reproduce using the example provided.

The issue does not seem to come from giotto-tda but rather from UMAP. Indeed, the UMAP class constructor has a random_state parameter, which is by default set to None, meaning that a "random random seed" is generated every time fit is called. Thus, running the UMAP part of the Mapper pipeline gives different results every time.

To have a fully reproducible pipeline, and solve your issue, set random_state in UMAP to some integer value of your choice.

[jnukpezah]

Hi @ulupo Got it! Thanks for the insight. Will do so! Great library btw!

[ulupo]

Great! Happy to help!