Stellargraph Versions Save

Fixed bugs:

• Fixed the issue in the APPNP class that causes appnp to propagate excessive dropout layers. #525
• Added a fix into the PPNP node classification demo so that the softmax layer is no longer propagated. #525

Fixed bugs:

• Updated requirements to Tensorflow>=1.14, as tensorflow with lower versions causes errors with sparse full batch node methods: GCN, APPNP, and GAT. #519

Fixed bugs:

• Reverted erroneous demo notebooks.

We are excited to announce the 0.8.0 release of the library. This release extends stellargraph by adding new algorithms and demos, enhancing interpretability via saliency maps for GAT, and further simplifying graph ML workflows through standardised model APIs and arguments. More details on new features and enhancements are listed below.

New algorithms:

• Directed GraphSAGE algorithm (a generalisation of GraphSAGE to directed graphs) + demo #479
• Attri2vec algorithm + demo #470 #455
• PPNP and APPNP algorithms + demos #485
• GAT saliency maps for interpreting node classification with Graph Attention Networks + demo #435

Implemented enhancements:

• New demo of node classification on Twitter hateful users \430
• New demo of graph saliency on Twitter hateful users #448
• Added Directed SampledBFS walks on directed graphs #464
• Unified API of GCN, GAT, GraphSAGE, and HinSAGE classses by adding build() method to GCN and GAT classes #439
• Added activations argument to GraphSAGE and HinSAGE classes #381
• Unified activations for GraphSAGE, HinSAGE, GCN and GAT #493 #381
• Added optional regularisation on the weights for GCN, GraphSage, and HinSage #172 #469
• Unified regularisation of GraphSAGE, HinSAGE, GCN and GAT #494 (geoffj-d61)
• Unsupervised GraphSage speed up via multithreading #474 #477
• Support of sparse generators in the GCN saliency map implementation. #432

Refactoring:

• Refactored Ensemble class into Ensemble and BaggingEnsemble. The former implements naive ensembles and the latter bagging ensembles. #459
• Changed from using keras to use tensorflow.keras #471
• Removed flatten_output arguments for all models #447

Fixed bugs:

• Updated Yelp example to support new dataset version #442
• Fixed bug where some nodes and edges did not get a default type #451
• Inconsistency in Ensemble.fit_generator() argument #461
• Fixed source--target node designations for code using Cora dataset #444
• IndexError: index 1 is out of bounds for axis 1 with size 1 in: demos/node-classification/hinsage #434
• GraphSAGE and GAT/GCN predictions have different shapes #425

Limited NetworkX version to <2.4 and Tensorflow version to <1.15 in installation requirements, to avoid errors due to API changes in the recent versions of NetworkX and Tensorflow.

Limited Keras version to <2.2.5 and Tensorflow version to <2.0 in installation requirements, to avoid errors due to API changes in the recent versions of Keras and Tensorflow.

Fixed bugs:

• Removed igraph and mplleaflet from demos requirements in setup.py. Python-igraph doesn't install on many systems and is only required for the clustering notebook. See the README.md in that directory for requirements and installation directions.
• Updated GCN interpretability notebook to work with new FullBatchGenerator API #429

New features and enhancements:

• SGC Implementation #361
• Updated to support Python 3.7 #348
• FullBatchNodeGenerator now supports a simpler interface to apply different adjacency matrix pre-processing options #405
• Full-batch models (GCN, GAT, and SGC) now return predictions for only those nodes provided to the generator in the same order #417
• GAT now supports using a sparse adjacency matrix making execution faster #420
• Added interpretability of GCN models and a demo of finding important edges for a node prediction #383
• Added a demo showing inductive classification with the PubMed dataset #372

Refactoring:

• Added build() method for GraphSAGE and HinSAGE model classes #385 This replaces the node_model() and link_model() methods, which will be deprecated in future versions (deprecation warnings added).
• Changed the FullBatchNodeGenerator to accept simpler method and transform arguments #405

Fixed bugs:

• Removed label from features for pubmed dataset. #362
• Python igraph requirement fixed #392
• Simplified random walks to not require passing a graph #408

Fixed bugs:

• a bug in passing graph adjacency matrix to the optional func_opt function in FullBatchNodeGenerator class
• a bug in demos/node-classification/gcn/gcn-cora-example.py:144: incorrect argument was used to pass the optional function to the generator for GCN

Enhancements:

• separate treatment of gcn and gat models in demos/ensembles/ensemble-node-classification-example.ipynb

New features and enhancements:

• Graph Attention (GAT) layer and model (stack of GAT layers), with demos #216, #315
• Unsupervised GraphSAGE #331 with a demo #335
• Model Ensembles #343
• Community detection based on unsupervised graph representation learning #354
• Saliency maps and integrated gradients for model interpretability #345
• Shuffling of head nodes/edges in node and link generators at each epoch #298

Fixed bugs:

• a bug where seed was not passed to sampler in GraphSAGELinkGenerator constructor #337
• UniformRandomMetaPathWalk doesn't update the current node neighbors #340
• seed value for link mapper #336