search

medialab / ipysigma

A Jupyter widget using sigma.js to render interactive networks.
https://medialab.github.io/ipysigma/demo.html
MIT License
188 stars 16 forks source link

readme

DOI

ipysigma

A Jupyter widget using sigma.js and graphology to render interactive networks directly within the result of a notebook cell.

Live Demo

ipysigma has been designed to work with either networkx or igraph.

ipysigma lets you customize a large number of the graph's visual variables such as: node color, size, label, border, halo, pictogram, shape and edge color, size, type, label etc.

For an exhaustive list of what visual variables you may tweak, check the "Available visual variables" part of the documentation.

ipysigma

ipysigma is also able to display synchronized & interactive "small multiples" of a same graph to easily compare some of its features.

ipysigma-grid

Summary

Installation

You can install using pip:

pip install ipysigma

You will also need to install either networkx or igraph.

If you are using an older version of Jupyter, or if the extension does not appear to be installed automatically, you might also need to run some nbextension/labextension commands likewise:

# Try one/all of those for jupyter notebook:
jupyter nbextension enable --py --sys-prefix ipysigma
jupyter nbextension enable --py --user ipysigma
jupyter nbextension enable --py --system ipysigma

# Try one/all of those for jupyter lab:
jupyter labextension enable ipysigma
jupyter labextension enable ipysigma user
jupyter labextension enable ipysigma sys-prefix

If you want to use ipysigma on Google Colab, you will need to enable widget output using the following code:

from google.colab import output

output.enable_custom_widget_manager()

Remember you can always install packages in Colab by executing the following command in a cell:

!pip install networkx ipysigma

Quick start

Using networkx

import networkx as nx
from ipysigma import Sigma

# Importing a gexf graph
g = nx.read_gexf('./my-graph.gexf')

# Displaying the graph with a size mapped on degree and
# a color mapped on a categorical attribute of the nodes
Sigma(g, node_size=g.degree, node_color='category')

Using igraph

import igraph as ig
from ipysigma import Sigma

# Generating a graph
g = ig.Graph.Famous('Zachary')

# Displaying the graph with a size mapped on degree and
# a color mapped on node betweenness centrality, using
# a continuous color scale named "Viridis"
Sigma(g, node_size=g.degree, node_color=g.betweenness(), node_color_gradient='Viridis')

Examples

Compute a Louvain partition and use it as node color

ipysigma is able to compute metrics on the widget side using graphology. As such, you can ask it to compute e.g. a Louvain partitioning if you don't want or cannot do it on the python side.

For more information about available metrics and how to specify them, check this part of the documentation.

Sigma(g, node_metrics=["louvain"], node_color="louvain")

# Renaming the target attribute
Sigma(g, node_metrics={"community": "louvain"}, node_color="community")

# Passing custom parameters
Sigma(
  g,
  node_metrics={"community": {"name": "louvain", "resolution": 1.5}},
  node_color="community"
)

Use networkx/igraph/custom metrics as visual variables

Use networkx metrics:

import networkx as nx

g = nx.path_graph(5)
Sigma(g, node_size=nx.eigenvector_centrality(g))

Use igraph metrics:

import igraph as ig

g = ig.Graph.GRG(5, 0.5)
Sigma(g, node_size=g.pagerank(), node_color=g.connected_components())

Use custom metrics:

import networkx as nx

def even_or_odd(node):
  return node % 2 == 0

g = nx.path_graph(5)
Sigma(g, node_color=even_or_odd)

Read this for an exhaustive list of what can be used as visual variables.

Display a pandas DataFrame as a graph

Converting tabular data to a graph is not obvious. So for this, we advise to use helper functions found in our other library python pelote.

In this first example, we create a graph from a DataFrame of edges:

import pandas as pd
from pelote import edges_table_to_graph

# Alice invited Bob and Chloe. Bob invited Chloe twice.
df = pd.DataFrame({
  "hosts": ["Alice", "Alice", "Bob", "Bob"],
  "guests": ["Bob", "Chloe", "Chloe", "Chloe"]
})

g = edges_table_to_graph(
  df,
  edge_source_col="hosts",
  edge_target_col="guests",
  count_rows_as_weight=True,
  directed=True
)

Sigma(g, edge_size='weight')

Using pelote again, you can also create a bipartite network (students and their professors, for example) with the table_to_bipartite_graph function:

import pandas as pd
from pelote import table_to_bipartite_graph

df = pd.DataFrame({
  "professor": ["A", "A", "A", "B", "B", "B", "B"],
  "student": ["C", "D", "E", "C", "F", "G", "H"],
})

g = table_to_bipartite_graph(df, 'student', 'professor', node_part_attr='status')

Sigma(g, node_color='status', default_node_size=10, show_all_labels=True)

Comparing two features of a graph

Let's say we have a graph of websites that we categorized by type and language and we want to compare the distribution of those categories on the graph's topology. We could use node color for language and border color for type but you will quickly see that this is probably not readable.

To solve this kind of problems and enable its users to easily compare multiple features of a graph, ipysigma exposes a SigmaGrid widget that arranges multiple synchronized views of the same graph on a grid:

from ipysigma import SigmaGrid

# Views to display can be specified through the `views` kwarg, expecting
# a list of dicts of keyword arguments to give to the underlying Sigma widgets:
SigmaGrid(g, views=[
  {"node_color": "type"},
  {"node_color": "type"}
])

# You can do the same by using the `#.add` method of the grid to
# dynamically add views:
SigmaGrid(g).add(node_color="lang").add(node_color="type")

# Any kwarg passed to the grid directly will be used by all of the views.
# This is useful to avoid repetition:
SigmaGrid(g, node_size=g.degree, views=[
  {"node_color": "type"},
  {"node_color": "type"}
])

# You can of course display more than 2 views
# By default the grid has 2 columns and will wrap to new rows,
# but you can change the number of columns using the `columns` kwarg:
SigmaGrid(g, columns=3, views=[
  {"node_size": g.degree},
  {"node_size": g.in_degree},
  {"node_size": g.out_degree}
])

More examples: functional testing notebooks

If you want comprehensive examples of the widget's visual variables being used, you can read the notebooks found here, which serve as functional tests to the library.

What data can be used as visual variable

There are several ways to specify what you want to use as visual variables (read this for a detailed explanation).

Here is the exhaustive list of what is possible:

Name of a node or edge attribute

# Let's say your nodes have a "lang" attribute, we can use its modalities as values for
# a categorical color palette:
Sigma(g, node_color='lang')

Node or edge mapping

# You can store the data in a mapping, e.g. a dictionary, likewise:
node_lang = {'node1': 'en', 'node2': 'fr', ...}
Sigma(g, node_color=node_lang)

# For edges, the mapping's key must be a 2-tuple containing source & target nodes.
# Note that for undirected graphs, the order of nodes in the tuple
# does not make any difference as both will work.
edge_type = {('node1', 'node2'): 'LIKES', ('node2', 'node3'): 'LOVES'}

Arbitrary iterable

# Any arbitrary iterable such as generators, ranges, numpy vectors,
# pandas series etc. will work. The only requirement is that they should
# follow the order of iteration of nodes or edges in the graph, so we may
# align the data properly.

# Creating a 0 to n generic label for my nodes
Sigma(g, node_label=range(len(g)))

# Random size for my edges
Sigma(g, edge_size=(random() for _ in g.edges))

# Numpy vector
Sigma(g, node_size=np.random.rand(len(g)))

# Pandas series
Sigma(g, edge_size=df.edge_weights)

Partition

# A partition, complete or not, but not overlapping, of nodes or edges:
# Must be a list of lists or a list of sets.
communities = [{2, 3, 6}, {0, 1}, {4, 6}]

Sigma(g, node_color=communities)

networkx/igraph degree view

# Mapping node size on degree is as simple as:
Sigma(g, node_size=g.degree)

igraph VertexClustering

# IGraph community detection / clustering methods return a VertexClustering object
Sigma(g, node_color=g.connected_components())

Sigma(g, node_color=g.community_multilevel())

Arbitrary callable

# Creating a label for my nodes
Sigma(g, node_label=lambda node: 'Label of ' + str(node))

# Using edge weight as size only for some source nodes
Sigma(g, edge_size=lambda u, v, a: attr['weight'] if g.nodes[u]['part'] == 'main' else 1)

# Node callables will be given the following arguments:
#   1. node key
#   2. node attributes

# Edge callables will be given the following arguments:
#  1. source node key
#  2. target node key
#  3. edge attributes

# Note that given callables may choose to take any number of those arguments.
# For instance, the first example only uses the first argument but still works.

Set

# A set will be understood as a binary partition with nodes or edges being
# in it or outside it. This will be mapped to a boolean value, with `True`
# meaning the node or edge was in the partition.

# This will display the nodes 1, 5 and 6 in a color, and all the other ones
# in a different color.
Sigma(g, node_color={1, 5, 6})

Visual variables and kwargs naming rationale

ipysigma lets its users tweak a large number of visual variables. They all work through a similar variety of keyword arguments given to the Sigma widget.

In ipysigma visual variables can be given:

kwargs naming rationale

To be able to be drawn on screen, every visual variable must use values that have a meaning for the the widget's visual representation. For colors, it might be a HTML color name such as #fa65ea or cyan. For sizes, it might be a number of pixels etc.

If you know what you are doing and want to give ipysigma the same "raw" values as those expected by the visual representation directly, all variables have kwargs starting by raw_, such as raw_node_color.

But if you want ipysigma to map your arbitrary values to a suitable visual representation, all variables have a kwarg without any prefix, for instance node_color.

In which case, if you use categorical data, ipysigma can generate or use palettes to map the category values to e.g. colors on screen. You can always customize the palette or mapping using a kwarg suffixed with _palette or _mapping such as node_color_palette or node_shape_mapping.

And if you use numerical data, then values will be mapped to an output range, usually in pixels, that can be configured with a kwarg suffixed with _range such as node_size_range. Similarly, if you want to map numerical data to a gradient of colors, you will find kwarg suffixed with _gradient such as node_color_gradient.

Sometimes, some values might fall out of the represented domain, such as non-numerical values for continuous variables, or categories outside of the colors available in the given palette. In which case there always exists a kwarg prefixed with default_, such as default_node_color. A neat trick is also to use those kwargs as a way to indicate a constant value if you want all your edges to have the same color for instance, or your nodes to have the same size in pixels.

Finally, it's usually possible to tweak the way numerical values will be mapped from their original domain to the visual one. This is what you do, for instance, when you choose to use a logarithmic scale on a chart to better visualize a specific distribution. Similarly, relevant ipysigma visual variables give access to a kwarg suffixed _scale, such as node_color_scale that lets you easily switch from a linear to a logarithmic or power scale etc. (for more information about this, check this in the next part of the documentation).

To summarize, let's finish with two exhaustive examples: node color & node size.

Categorical or continuous variable: node color as an example

Continuous variable: node size as an example

For a comprehensive view of the available visual variables, the values they expect and how they can be customized, read this next part of the documentation.

Scales, palettes and gradients

Available scales

All the _scale kwargs can take the following:

Color palettes

By default, color palettes are generated for you by ipysigma using iwanthue. ipysigma will first count the number of distinct categories to represent, sort them by frequency and generate a palette of up to 10 colors for the most used ones. The other one will use the default one given to the relevant default_ kwarg such as default_node_color for instance.

Note that this maximum number of 10 can be increased using the max_categorical_colors kwarg.

Note also that the palette generation is seeded using the mapped attribute name in the data so that the palette is always the same (if the name and the category count remains the same), but is different from one attribute to the other.

If you don't want ipysigma to generate color palettes for you, you can give your own palette through the relevant _palette kwarg such as node_color_palette, or use some d3-scale-chromatic one (they have names starting with scheme).

Here is the full list of those palettes supported by ipysigma: Accent, Blues, BrBG, BuGn, BuPu, Category10, Dark2, GnBu, Greens, Greys, OrRd, Oranges, PRGn, Paired, Pastel1, Pastel2, PiYG, PuBu, PuBuGn, PuOr, PuRd, Purples, RdBu, RdGy, RdPu, RdYlBu, RdYlGn, Reds, Set1, Set2, Set3, Spectral, Tableau10, YlGn, YlGnBu, YlOrBr, YlOrRd.

Color gradients

Color gradients can be defined as a range from "lowest" to "highest" color, e.g. ("yellow", "red).

They can also be taken from any d3-scale-chromatic continuous gradient (they have names starting with interpolate).

Here is the full list of those gradients supported by ipysigma: Blues, BrBG, BuGn, BuPu, Cividis, Cool, CubehelixDefault, GnBu, Greens, Greys, Inferno, Magma, OrRd, Oranges, PRGn, PiYG, Plasma, PuBu, PuBuGn, PuOr, PuRd, Purples, Rainbow, RdBu, RdGy, RdPu, RdYlBu, RdYlGn, Reds, Sinebow, Spectral, Turbo, Viridis, Warm, YlGn, YlGnBu, YlOrBr, YlOrRd.

Widget-side metrics

Since ipysigma is using graphology, it can also draw from its library of graph theory metrics.

As such, the node_metrics enables you to ask your widget to compute node metrics on its own and use to map the result on any visual variable.

Here is how you can specify metrics to be computed:

# node_metrics expects an iterable of metrics to compute:
Sigma(g, node_metrics=["louvain"], node_color="louvain")

# They can be specified by name, but you can also specify through
# a dictionary if you need parameters for the metrics:
Sigma(g, node_metrics=[{"name": "louvain", "resolution": 1.5}], node_color="louvain")

# You can also give a dictionary mapping resulting attribute name to
# the metric to compute if you don't want to map the result on an attribute
# having the same name as the metric:
Sigma(g, node_metrics={"community": "louvain"}, node_color="community")
Sigma(g, node_metrics={"community": {"name": "louvain", "resolution": 1.5}}, node_color="community")

Available node metrics & their parameters

Frequently asked questions

Why are there so few labels displayed?

Labels are costly to render and can negate the benefit of using a WebGL renderer such as sigma.js to render interactive graphs. As such, sigma.js relies on a constant size grid to select the "worthiest" labels to display, after taking camera zoom into account.

You can tweak the parameters of this grid using label_grid_cell_size and label_density. Decreasing the first one or increasing the second one will result in more labels being displayed.

Also, by default, the label of a node is displayed only if its size in pixels is larger than a threshold. You can change that threshold using the label_rendered_size_threshold kwarg.

Finally, if you don't want to deal with all this nonsense and just want to display all labels because you know what you are doing and don't care about performance, you can just use show_all_labels=True instead.

Why are some of my categories mapped to a dull grey?

When ipysigma generates palettes for you, it only uses up to 10 colors by default. This number can be increased using the max_categorical_colors kwarg. For more information about palette generation, read this part of the documentation.

Some designer told me (while holding a baseball bat) that it is unwise to have more than 10 categorical colors because you won't be able to distinguish them anymore. My hands are tied. Don't ask me to change this.

I gave colors to node_color but arbitrary colors are displayed by the widget instead

node_color does not expect colors per se but arbitrary data that will be mapped to a suitable color palette for you. If you want to give colors directly, use raw_node_color instead. For more information about the visual variables kwarg naming rationale, read this part of the documentation.

My computer sounds like an airplane taking off

Don't forget to turn off the layout when it has converged (the pause button on the left). There is no convincing way to automatically detect when layout has converged so we must rely on you, the user, to indicate when it's done.

If you want to start the layout automatically when instantiating the widget and make sure it will automatically stop after, say, 10 seconds, use start_layout=10.

Some of my widgets only display labels or a glitchy black box

Your GPU can only render so many webgl canvases in your browser tabs. So if you created too many widgets (this depends on the specifics of your computer and graphics card), it may gracefully deal with the situation by erasing the graph (but not the labels since those are rendered using 2d canvases) or by glitching to death.

My graph is ugly, make it beautiful like Gephi

Use default_edge_type="curve", node_border_color_from="node", label_size=g.degree and label_font="cursive" and you should have a dazzling Gephi graph.

gephi

Available visual variables

node_color

node_color

Type

Categorical or continuous.

Raw values

HTML named color or hex color or rgb/rgba color. Examples: red, #fff, #a89971, rgb(25, 25, 25), rgba(25, 145, 56, 0.5)

Related kwargs

node_color_saturation

node_color_saturation

Type

Continuous.

Raw values

A percentage of color saturation. Examples: 0.1, 0.96.

Related kwargs

node_size

node_size

Type

Continuous.

Raw values

A node size, i.e. a circle radius, in pixels, with default camera (not zoomed nor unzoomed).

Related kwargs

node_label

node_label

Type

Raw only.

Raw values

A text label.

Related kwargs

node_label_size

node_label_size

Type

Continuous.

Raw values

A font size for the label text, in pixels.

Related kwargs

node_label_color

node_label_color

Type

Categorical.

Raw values

HTML named color or hex color or rgb/rgba color. Examples: red, #fff, #a89971, rgb(25, 25, 25), rgba(25, 145, 56, 0.5)

Related kwargs

node_border_size

node_border_size

Type

Continuous.

Raw values

A border size, in pixels, with default camera (not zoomed nor unzoomed).

Note that this border size will be added to the node's radius.

Related kwargs

Notes

Borders are only shown on screen if a node_border_size OR a node_border_ratio AND a node_border_color are defined.

node_border_ratio

node_border_ratio

Type

Continuous.

Raw values

A border ratio, in percentage, with default camera (not zoomed nor unzoomed).

Note that this border ratio will eat the node's size.

Related kwargs

Notes

Borders are only shown on screen if a node_border_size OR a node_border_ratio AND a node_border_color are defined.

node_border_color

node_border_color

Type

Categorical or continuous.

Raw values

HTML named color or hex color or rgb/rgba color. Examples: red, #fff, #a89971, rgb(25, 25, 25), rgba(25, 145, 56, 0.5)

Related kwargs

Notes

Borders are only shown on screen if a node_border_size OR a node_border_ratio AND a node_border_color are defined.

node_pictogram

node_pictogram

Type

Categorical.

Raw values

The name of any Google Material Icon as listed here (the name must be lowercase and snake_case, e.g. the name "Arrow Drop Done" should be given to ipysigma as arrow_drop_done).

Alternatively, one can also give urls of publicly accessible svg icons such as https://fonts.gstatic.com/s/i/short-term/release/materialsymbolsoutlined/arrow_drop_down/default/48px.svg

Related kwargs

Notes

Pictograms are only shown on screen if node_pictogram AND node_pictogram_color are defined.

node_pictogram_color

node_pictogram_color

Type

Categorical.

Raw values

HTML named color or hex color or rgb/rgba color. Examples: red, #fff, #a89971, rgb(25, 25, 25), rgba(25, 145, 56, 0.5)

Related kwargs

Notes

Pictograms are only shown on screen if node_pictogram AND node_pictogram_color are defined.

node_shape

node_shape

Type

Categorical.

Raw values

The name of a supported shape such as: circle, triangle, square, pentagon, star, hexagon, heart or cloud.

Alternatively, if you are feeling adventurous, it can also be the name of any Google Material Icon as listed here (the name must be lowercase and snake_case, e.g. the name "Arrow Drop Done" should be given to ipysigma as arrow_drop_done).

Finally, one can also give urls of publicly accessible svg icons such as https://fonts.gstatic.com/s/i/short-term/release/materialsymbolsoutlined/arrow_drop_down/default/48px.svg

Related kwargs

Note

Node shapes cannot be used with borders nor pictograms nor halos, as of yet.

node_halo_size

node_halo_size

Type

Continuous.

Raw values

A halo size offset in pixels, with default camera (not zoomed nor unzoomed). The full halo radius will therefore be its size + its node's radius.

Related kwargs

node_halo_color

node_halo_color

Type

Categorical or continuous.

Raw values

HTML named color or hex color or rgb/rgba color. Examples: red, #fff, #a89971, rgb(25, 25, 25), rgba(25, 145, 56, 0.5)

Related kwargs

edge_color

edge_color

Type

Categorical or continuous.

Raw values

HTML named color or hex color or rgb/rgba color. Examples: red, #fff, #a89971, rgb(25, 25, 25), rgba(25, 145, 56, 0.5)

Related kwargs

edge_type

edge_type

edge_size

edge_size

Type

Continuous.

Raw values

An edge thickness in pixels, with default camera (not zoomed nor unzoomed).

Related kwargs

edge_curveness

edge_curveness

Type

Continuous.

Raw values

A percentage. Note that it can go beyond 1 and that 0 will make the edge disappear.

Related kwargs

edge_label

edge_label

Type

Raw only.

Raw values

A text label.

Related kwargs

API Reference

Sigma

Arguments

.get_layout

Method returning the layout of the graph, i.e. the current node positions in the widget, as a dict mapping nodes to their {x, y} coordinates.

.get_camera_state

Method returning the current camera state of the widget, as a {x, y, ratio, angle} dict.

.get_selected_node

Method returning the currently selected node if any or None.

.get_selected_edge

Method returning the currently selected edge as a (source, target) tuple if any or None.

.get_selected_node_category_values

Method returning a set of currently selected node category values or None.

.get_selected_edge_category_values

Method returning a set of currently selected edge category values or None.

.render_snapshot

Method rendering the widget as an rasterized image in the resulting cell.

.to_html

Method rendering the widget as a standalone HTML file that can be hosted statically elsewhere.

Arguments

Sigma.set_defaults

Static method that can be used to override some default values of the Sigma class kwargs.

Arguments

Sigma.write_html

Static method taking the same kwargs as Sigma and rendering the widget as a standalone HTML file that can be hosted statically elsewhere.

Arguments

SigmaGrid

Arguments

.add

Method one can use as an alternative or combined to SigmaGrid constructor's views kwarg to add a new Sigma view to the grid. It takes any argument taken by Sigma and returns self for easy chaining.

SigmaGrid(g, node_color='category').add(node_size=g.degree).add(node_size='occurrences')

How to cite

Guillaume Plique. (2022). ipysigma, A Jupyter widget using sigma.js to render interactive networks. Zenodo. https://doi.org/10.5281/zenodo.7446059