Skip to content

@kjun9 kjun9 released this Jan 29, 2020 · 1976 commits to develop since this release

Major features and improvements

  • StellarGraph is now available as a conda package on Anaconda Cloud #516
  • New algorithms:
    • Cluster-GCN: an extension of GCN that can be trained using SGD, with demo #487
    • Relational-GCN (RGCN): a generalisation of GCN to relational/multi edge type graphs, with demo #490
    • Link prediction for full-batch models: FullBatchLinkGenerator allows doing link prediction with algorithms like GCN, GAT, APPNP and PPNP #543
  • Unsupervised GraphSAGE has now been updated and tested for reproducibility. Ensuring all seeds are set, running the same pipeline should give reproducible embeddings. #620
  • A datasets subpackage provides easier access to sample datasets with inbuilt downloading. #690

Breaking changes

  • The stellargraph library now only supports tensorflow version 2.0 #518, #732. Backward compatibility with earlier versions of tensorflow is not guaranteed.
  • The stellargraph library now only supports Python versions 3.6 and above #641. Backward compatibility with earlier versions of Python is not guaranteed.
  • The StellarGraph class no longer exposes NetworkX internals, only required functionality. In particular, calls like list(G) will no longer return a list of nodes; use G.nodes() instead. #297 If NetworkX functionality is required, use the new .to_networkx() method to convert to a normal networkx.MultiGraph or networkx.MultiDiGraph.
  • Passing a NodeSequence or LinkSequence object to GraphSAGE and HinSAGE classes is now deprecated and no longer supported #498. Users might need to update their calls of GraphSAGE and HinSAGE classes by passing generator objects instead of generator.flow() objects.
  • Various methods on StellarGraph have been renamed to be more succinct and uniform:
    • get_feature_for_nodes is now node_features
    • type_for_node is now node_type
  • Neighbourhood methods in StellarGraph class (neighbors, in_nodes, out_nodes) now return a list of neighbours instead of a set. This addresses #653. This means multi-edges are no longer collapsed into one in the return value. There will be an implicit change in behaviour for explorer classes used for algorithms like GraphSAGE, Node2Vec, since a neighbour connected via multiple edges will now be more likely to be sampled. If this doesn't sound like the desired behaviour, consider pruning the graph of multi-edges before running the algorithm.
  • GraphSchema has been simplified to remove type look-ups for individual nodes and edges #702 #703. Migration: for nodes, use StellarGraph.node_type; for edges, use the triple argument to the edges method, or filter when doing neighbour queries using the edge_types argument.
  • NodeAttributeSpecification and the supporting Converter classes have been removed #707. Migration: use the more powerful and flexible preprocessing tools from pandas and sklearn (see the linked PR for specifics)

Experimental features

Some new algorithms and features are still under active development, and are available as an experimental preview. However, they may not be easy to use: their documentation or testing may be incomplete, and they may change dramatically from release to release. The experimental status is noted in the documentation and at runtime via prominent warnings.

  • The StellarGraph and StellarDiGraph classes supports using a backend based on NumPy and Pandas that uses dramatically less memory for large graphs than the existing NetworkX-based backend #668. The new backend can be enabled by constructing with StellarGraph(nodes=..., edges=...) using Pandas DataFrames, instead of a NetworkX graph.

Bug fixes and other changes

  • Documentation for every relased version is published under a permanent URL, in addition to the stable alias for the latest release, e.g. https://stellargraph.readthedocs.io/en/v0.8.4/ for v0.8.4 #612
  • Neighbourhood methods in StellarGraph class (neighbors, in_nodes, out_nodes) now support additional parameters to include edge weights in the results or filter by a set of edge types. #646
  • Changed GraphSAGE and HinSAGE class API to accept generator objects the same as GCN/GAT models. Passing a NodeSequence or LinkSequence object is now deprecated. #498
  • SampledBreadthFirstWalk, SampledHeterogeneousBreadthFirstWalk and DirectedBreadthFirstNeighbours have been made 1.2-1.5× faster #628
  • UniformRandomWalk has been made 2× faster #625
  • FullBatchNodeGenerator.flow has been reduced from O(n^2) quadratic complexity to O(n), where n is the number of nodes in the graph, making it orders of magnitude faster for large graphs #513
  • The dependencies required for demos and testing have been included as "extras" in the main package: demos and igraph for demos, and test for testing. For example, pip install stellargraph[demos,igraph] will install the dependencies required to run every demo. #661
  • The StellarGraph and StellarDiGraph constructors now list their arguments explicitly for clearer documentation (rather than using *arg and **kwargs splats) #659
  • sys.exit(0) is no longer called on failure in load_dataset_BlogCatalog3 #648
  • Warnings are printed using the Python warnings module #583
  • Numerous DevOps changes:
Assets 2
You can’t perform that action at this time.