graphrole 1.1.1

Automatic feature extraction and node role assignment for transfer learning on graphs

GraphRole

Automatic feature extraction and node role assignment for transfer learning on graphs; based on the ReFeX/RolX algorithms [1, 2] of Henderson, et al.

Overview

A fundamental problem for learning on graphs is extracting meaningful features. GraphRole provides the RecursiveFeatureExtractor class to automate this process by extracting recursive features capturing local and neighborhood ("regional") structural properties of a given graph. The specific implementation follows that of the ReFeX algorithm [1]. Node features (e.g., degree) and ego-net features (e.g., neighbors, number of internal vs. external edges) are extracted and then recursively aggregated over each node's neighbors' features until no additional information is encoded. As is shown in [1], these recursive, "regional" features facilitate node classification and perform quite well in transfer learning tasks.

GraphRole also provides the RoleExtractor class for node role assignment (a form of classification). Different nodes play different structural roles in a graph, and using recursive regional features, these roles can be identified and assigned to collections of nodes. As they are structural in nature, node roles differ from and are often more intuitive than the commonly used communities of nodes. In particular, roles can generalize across graphs whereas the notion of communities cannot [2]. Identification and assignment of node roles has been shown to facilitate many graph learning tasks.

Please see [1, 2] for more technical details.

Installation

This package is hosted on PyPI and can be installed via pip:

$ pip install graphrole

To instead install from source:

$ git clone https://github.com/dkaslovsky/GraphRole.git
$ cd GraphRole
$ python setup.py install

Example

An example of GraphRole usage is found in the examples directory. The notebook example.ipynb (also available via nbviewer) walks through feature extraction and role assignment for the well-known karate_club_graph that is included with NetworkX. Recursive features are extracted and used to learn role assignments for each node in the graph. The graph is shown above with each node colored corresponding to its role.

The extracted roles reflect structural properties of the graph at the node level. The nodes 0 and 33 (dark green) are central to the graph and are connected to many other nodes. Nodes 1, 2, 3, and 32 are assigned to a similar role (red). In contrast, the roles colored as dark blue, light blue, and pink are found at the periphery of the graph. Notably, nodes need not be near one another to be assigned to the same role; instead nodes with similar properties are grouped together across the graph by their role assignments.

Although not reflected by this example, weighted and directed graphs are also supported and will yield weighted and directed variants of the extracted features.

Usage

For general usage, begin by importing the two feature and role extraction classes:

>>> from graphrole import RecursiveFeatureExtractor, RoleExtractor

Features are then extracted from a graph G into a pandas.DataFrame:

>>> feature_extractor = RecursiveFeatureExtractor(G)
>>> features = feature_extractor.extract_features()

Next, these features are used to learn roles. The number of roles is automatically determined by a model selection procedure when n_roles=None is passed to the RoleExtractor class instance. Alternatively, n_roles can be set to a desired number of roles to be extracted.

>>> role_extractor = RoleExtractor(n_roles=None)
>>> role_extractor.extract_role_factors(features)

The role assignment for each node can be retrieved as a dictionary:

>>> role_extractor.roles

Alternatively, roles can be viewed as a soft assignment and a node's percent membership to each role can be retrieved as a pandas.DataFrame:

>>> role_extractor.role_percentage

Node Attributes as User-Defined Features

GraphRole uses predefined structural graph properties for constructing features. It is also possible, as of version 1.1.0, to optionally include numeric node attributes as features. Providing a graph annotated with node attributes to GraphRole allows a user to seed the recursive feature extraction process with user-defined features for each node.

Node attributes are enabled by passing attributes=True as a kwarg to the RecursiveFeatureExtractor:

>>> feature_extractor = RecursiveFeatureExtractor(G, attributes=True)

Providing this kwarg will automatically include all numeric node attributes as features to be included in the recursive feature calculations. Attributes with non-numeric values are always skipped (set to zero). Note that the feature names associated with node attributes will be the provided attribute name prepended with a prefix of attribute_.

A list of attributes to be included for feature calculation instead of defaulting to all numeric attributes can be provided as:

>>> feature_extractor = RecursiveFeatureExtractor(G, attributes=True, attributes_include=['attr1', 'attr3'])

which will specify the use of only node attributes attr1 and attr3.

A list of attributes to be excluded for feature calculation from the default of all numeric attributes can be provided as:

>>> feature_extractor = RecursiveFeatureExtractor(G, attributes=True, attributes_exclude=['attr2'])

which will specify the use of all node attributes other than attr2.

For safety, the attributes_exclude list takes priority over the attributes_include list when conflicting specifications are provided.

Note: igraph uses the attribute name to store an identifier for all nodes and therefore the corresponding attribute value is never used for feature calculations. The attribute name, even if overwritten by the user, is always skipped for igraph graph instances.

Graph Interfaces

An interface for graph data structures is provided in the graphrole.graph.interface module. Implementations for networkx and igraph are included.

The igraph package is not included in requirements.txt and thus will need to be manually installed if desired. This is due to additional installation requirements beyond pip install python-igraph; see the igraph documentation for more details. Note that all tests that require igraph will be skipped if it is not installed.

To add an implementation of an additional graph library or data structure:

1. Subclass the BaseGraphInterface ABC in graphrole.graph.interface.base.py and implement the required methods
2. Update the INTERFACES dict in graphrole.graph.interface.__init__.py to make the new subclass discoverable
3. Add tests by trivially implementing a setUpClass() classmethod of a subclass of BaseGraphInterfaceTest.BaseGraphInterfaceTestCases in the tests.test_graph.test_interface.py module
4. If desired, a similar procedure allows the feature extraction tests to be run using the added interface by again trivially implementing a setUpClass() classmethod of a subclass of BaseRecursiveFeatureExtractorTest.TestCases in the tests.test_features.test_extract.py module

Future Development

Model explanation ("sense making") will be added to the RoleExtractor class in a future release.

Tests

To run tests:

$ python -m unittest discover -v

As noted above, the tests for the igraph interface are skipped when igraph is not installed. Because this package is intentionally not required, the test coverage reported above is much lower than when igraph is installed and its interface tests are not skipped (97% coverage to date).

References

[1] Henderson, et al. It’s Who You Know: Graph Mining Using Recursive Structural Features.

[2] Henderson, et al. RolX: Structural Role Extraction & Mining in Large Graphs.