alph - altair for your graph

A Python library using Altair for declarative, data-driven network visualisation.

Why

Tidy, legible graph visualisations can be elusive. Alph helps by bringing together effective styling, layout and pruning options from across the Python ecosystem.

How it works

1. Get your data into a NetworkX Graph
2. Pick a network layout function, or bring your own node coordinates
3. Define node and edge style attributes
4. Plot using a simple function call

Best bet is probably to dive straight into the examples, and come back to the API documentation below as needed.

Features

• plot any NetworkX Graph
• support for any layout expressed as a NetworkX pos structure (a dict like {node_id: (x,y), ...})
• several readily accessible and tuneable layout functions (see example)
• Altair-style data driven node and edge styling - size, colour, stroke, opacity, scales, conditionals and more
• convenience args for node labels, halos
• experimental 1-level "combo" node support

Installation

Minimal

pip install alph

Recommended

1. for graphviz layout support, install graphviz on your platform - e.g. brew install graphviz on Mac, sudo apt install libgraphviz-dev graphviz on Colab, Debian, Ubuntu etc - or download the installer

2. Install alph with graphviz support:

   pip install "alph[graphviz]"
   

3. Install the ForceAtlas2 layout library from our fork, along with cython for speedup

   pip install cython "git+https://github.com/connectedcompany/forceatlas2.git"
   

Why is the ForceAtlas2 install separate?

ForceAtlas2 is a classic, user feedback led layout algorithm from the Gephi team, and the ForceAtlas2 library implementation is an excellent, performant Python port. There are two things to be aware of:

1. GPL licence - the ForceAtlas2 library, and some of the works it is derived from, are GPL licensed - hence care is needed when distributing and linking to it. We thus intend to keep its install optional long term. Since alph uses a plugin design for layout providers, this is straightforward for us to accommodate, and maintain explicit separation for use cases where GPL is an issue.

2. Our fork - recently, releases of the library have been sporadic - though there is stated intent for regular maintenance to resume. In the meantime, we've created a temporary fork to be able to roll in changes. Currently, our fork incorporates a simple change that enables deterministic layouts.

Usage

Simply call the alph function with desired options.

Minimally, given a weighted network G:

from alph import alph

G = ...
alph(G, weight_attr="weight")

Examples

See examples. Here's an overview:

• Some of the supported layouts (from the layouts gallery example):

• Use of geolocation coordinates for graph layout, (from the flight routes example):

• Basic styling (from the styling example):

• A styled interaction network (from the dolphins example):

• "Combo"-style layout (experimental) - category-driven node grouping with edge weight aggregation, from the Les Mis example:

---

API

Supported layout functions

• NetworkX layouts: Spring, Fruchterman-Reingold, etc
• NetworkX-wrapped graphviz layouts: dot, neato etc. You can use args for specific layouts - for example, for neato: args="-Goverlap=scale -Gstart=123"
• Gephi ForceAtlas2 based on the forceatlas2 Python implementation - see layout.py for configuration options, and this paper for more detail
• ForceAtlas implementation within scikit-network
• Any other that returns a NetworkX-style node positions dictionary

Supported arguments

arg	type(s)	default	description
G	Networkx Graph		graph to visualise
weight_attr	str		edge weight attribute, for weighted graphs
layout_fn	function	ForceAtlas2	Function that, given a graph, returns a layout
node args	dict		See below
edge args	dict		See below
combo_group_by	str or list		Attribute to use to create grouped combo nodes
combo_layout_fn	function	Fruchterman-Reingold	Layout function for combo nodes
combo_node_args	dict		See below
combo_edge_args	dict		See below
combo_edge_weight_agg_attr	dict		Attribute to use to weigh combo edges; if set, overrides weight_attr. Can use values given via combo_edge_agg_attrs. If not set and weight_attr not given, falls back to simple edge count.
combo_edge_agg_attrs	dict		Pandas groupby-style dict, describing how to aggregate edge attributes that span nodes - for example {"combo_edge_attr_name": ("edge_attr_name", "sum")}
combo_edge_weight_threshold	dict		Drop edges below this weight
include_edgeless_combo_nodes	dict		Whether or not to incorporate disconnected combo nodes
combo_node_additional_attrs	dict		Attributes to add to combo nodes
edge_node_additional_attrs	dict		Attributes to add to combo node edges, like {"edge_attr_name": agg_fn}; agg fn is applied across all attr values for edges that link grouped nodes
combo_empty_attr_action	drop, group or promote	drop	What to do with nodes that have an empty value for the combo_group_by attribute
combo_size_scale_domain	2-item list or tuple	[0, 25]	Lower/upper bound of num nodes to apply to combo node size range
combo_size_scale_range	2-item list or tuple	[6**2, 180**2]	Combo node size range
combo_inner_graph_scale_factor	float	0.6	Scale down inner graph to fit inside combo nodes by this factor - normally <1
non_serializable_datetime_format	str	%d %b %Y	Format string for non-serialisable date / time types that otherwise break Altair
width	int	800	Figure width (px)
height	int	600	Figure height (px)
prop_kwargs	dict		Optional properties such as title
padding	int		Padding inside figure edges. No node centres will be placed outside this boundary.
nodes_layer_params	selection or other		Altair params to be added to the nodes layer via .add_params() - typically a selection

Node args

arg	type(s)	default
size	int, alt.*	400
tooltip_attrs	list
fill	str, alt.*	#000
opacity	float, alt.*	1
stroke	str, alt.*	#565656
strokeWidth	int, alt.*	0
halo_offset	int
halo_opacity	float, alt.*	1
halo_stroke	str, alt.*	#343434
halo_strokeWidth	int, alt.*	2
label_attr	str
label_offset	int	6
label_size	int	10
label_color	str	black

Edge args

arg	type(s)	default
weight_attr	str	weight
color	str, alt.*	#606060
opacity	float , alt.*	alt.Size(weight_attr, scale=alt.Scale(range=[0.3, 1]) if weighted, 1 otherwise
strokeWidth	int, alt.*	alt.Size(weight_attr, scale=alt.Scale(range=[0.1, 5]) if weighted, 2 otherwise

---

Tips

• Get to know your layout algos - especially the 2-3 most used ones. They can have a dramatic effect on the results. A combination of FA2, Spring and Fruchterman is extremely versatile if set up right.
• Pass seed to layout functions where possible, for repeatable layouts
• Set padding to ensure node elements stay within figure boundaries

Known limitations

• Node size attribute does not support all Altair options - currently only alt.value and alt.Size with linear domain and range scales. More can be added as needed.

  This is a design choice, made to not burden the user with calculating things like label and halo positions when node sizes vary. Will review this tradeoff based on in-use experience.

• One combo level currently supported

See also

• nx-altair is a nice project that takes a slightly different approach to combining NetworkX and Altair for network viz.