Skip to content

Latest commit

 

History

History
656 lines (503 loc) · 89.9 KB

CHANGELOG.md

File metadata and controls

656 lines (503 loc) · 89.9 KB
Raw

Change Log

1.2.1

Full Changelog

This release is a small bug fix release on top of 1.2.0.

Bug fixes and other changes

  • Update the URLs of some datasets (Cora, PubMedDiabetes, CiteSeer) for upstream changes #1738, #1759
  • Add two missing layers to the stellargraph.custom_keras_layers dictionary #1757
  • Experimental changes: rename RotHEScoring to RotHEScore #1756
  • DevOps:

1.2.0

Full Changelog

Jump in to this release, with the new and improved demos and examples:

Major features and improvements

  • Better Windows support: StellarGraph's existing ability to run on Windows has been improved, with all tests running on CI (#1696) and several small fixes (#1671, #1704, #1705).
  • Edge weights are supported in GraphSAGE (#1667) and Watch Your Step (#1604). This is in addition to the existing support for edge weights in GCN, GAT, APPNP, PPNP, RGCN, GCN graph classification, DeepGraphCNN and Node2Vec sampling.
  • Better and more demonstration notebooks and documentation to make the library more accessible to new and existing users:
    • A demo notebook for a comparison of link prediction with random walks based node embedding, showing Node2Vec, Attri2Vec, GraphSAGE and GCN #1658
    • The demo notebook for unsupervised training with Deep Graph Infomax has been expanded with more explanation and links #1257
    • The documentation for models, generators and other elements now has many more links to other relevant items in a "See also" box, making it easier to fit pieces together (examples: GraphSAGE, GraphSAGENodeGenerator, BiasedRandomWalk) #1718
  • The Cluster-GCN training procedure supports unsupervised training via Deep Graph Infomax; this allows for scalable training of GCN, APPNP and GAT models, and includes connecting to Neo4j for large graphs demo (#1257)
  • KGTripleGenerator now supports the self-adversarial negative sampling training procedure for knowledge graph algorithms (from RotatE), via generator.flow(..., sample_strategy="self-adversarial") docs

Deprecations

  • The ClusterGCN model has been replaced with the GCN class. In the previous 1.1.0 release, GCN, APPNP and GAT were generalised to support the Cluster-GCN training procedure via ClusterNodeGenerator (which includes Neo4j support). The ClusterGCN model is now redundant and thus is deprecated: however, it still works without behaviour change.

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.

  • RotE, RotH: knowledge graph link prediction algorithms that combine TransE and RotatE in Euclidean or hyperbolic space, respectively #1539

Bug fixes and other changes

1.1.0

Full Changelog

Jump in to this release, with the new and improved demos and examples:

Major features and improvements

  • Support for the Neo4j graph database has been significantly improved:
    • There is now a Neo4jStellarGraph class that packages up a connection to a Neo4j instance, and allows it to be used for machine learning algorithms including the existing Neo4j and GraphSAGE functionality demo, #1595, #1598.
    • The ClusterNodeGenerator class now supports Neo4jStellarGraph in addition to the in-memory StellarGraph class, allowing it to be used to train models like GCN and GAT with data stored entirely in Neo4j demo (#1561, #1594, #1613)
  • Better and more demonstration notebooks and documentation to make the library more accessible to new and existing users:
  • New algorithms:
  • The StellarGraph class continues to get smaller, faster and more flexible:
    • Node features can now be specified as NumPy arrays or the newly added thin IndexedArray wrapper, which does no copies and has minimal runtime overhead demo (#1535, #1556, #1599). They can also now be multidimensional for each node #1561.
    • Edges can now have features, taken as any extra/unused columns in the input DataFrames demo #1574
    • Adjacency lists used for random walks and GraphSAGE/HinSAGE are constructed with NumPy and stored as contiguous arrays instead of dictionaries, cutting the time and memory or construction by an order of magnitude #1296
    • The peak memory usage of construction and adjacency list building is now monitored to ensure that there are not large spikes for large graphs, that exceed available memory #1546. This peak usage has thus been optimised: #1551,
    • Other optimisations: the edge_arrays, neighbor_arrays, in_node_arrays and out_node_arrays methods have been added, reducing time and memory overhead by leaving data as its underlying NumPy array #1253; the node_type method now supports multiple nodes as input, making algorithms like HinSAGE and Metapath2Vec much faster #1452; the default edge weight of 1 no longer consumes significant memory #1610.
  • Overall performance and memory usage improvements since 1.0.0, in numbers:
    • A reddit graph has 233 thousand nodes and 11.6 million edges:
      • construction without node features is now 2.3× faster, uses 31% less memory and has a memory peak 57% smaller.
      • construction with node features from NumPy arrays is 6.8× faster, uses 6.5% less memory overall and 85% less new memory (the majority of the memory is shared with the original NumPy arrays), and has a memory peak (above the raw data set) 70% smaller, compared to Pandas DataFrames in 1.0.0.
      • adjacency lists are 4.7-5.0× faster to construct, use 28% less memory and have a memory peak 60% smaller.
    • Various random walkers are faster: BiasedRandomWalk is up to 30× faster with weights and 5× faster without weights on MovieLens and up to 100× faster on some synthetic datasets, UniformRandomMetapathWalk is up to 17× faster (on MovieLens), UniformRandomWalk is up to 1.4× (on MovieLens).
  • TensorFlow 2.2 and thus Python 3.8 are now supported #1278

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.

  • RotatE: a knowledge graph link prediction algorithm that uses complex rotations (|z| = 1) to encode relations #1522
  • GCN_LSTM (renamed from GraphConvolutionLSTM): time series prediction on spatio-temporal data. It is still experimental, but has been improved since last release:
    • the SlidingFeaturesNodeGenerator class has been added to yield data appropriate for the model, straight from a StellarGraph instance containing time series data as node features #1564
    • the hidden graph convolution layers can now have a custom output size #1555
    • the model now supports multivariate input and output, including via the SlidingFeaturesNodeGenerator class (with multidimensional node features) #1580
    • unit tests have been added #1560
  • Neo4j support: some classes have been renamed from Neo4J... (uppercase J) to Neo4j... (lowercase j).

Bug fixes and other changes

  • Edge weights are supported in methods using FullBatchNodeGenerator (GCN, GAT, APPNP, PPNP), RelationalFullBatchNodeGenerator (RGCN) and PaddedGraphGenerator (GCN graph classification, DeepGraphCNN), via the weighted=True parameter #1600
  • The StellarGraph class now supports conversion between node type and edge type names and equivalent ilocs #1366, which allows optimising some algorithms (#1367 optimises ranking with the DistMult algorithm from 42.6s to 20.7s on the FB15k dataset)
  • EdgeSplitter no longer prints progress updates #1619
  • The info method now merges edge types triples like A-[r]->B and B-[r]->A in undirected graphs #1650
  • There is now a notebook capturing time and memory resource usage on non-synthetic datasets, designed to help StellarGraph contributors understand and optimise the StellarGraph class #1547
  • Various documentation, demo and error message fixes and improvements: #1516 (thanks @thatlittleboy), #1519, #1520, #1537, #1541, #1542, #1577, #1605, #1606, #1608, #1624, #1628, #1632, #1634, #1636, #1643, #1645, #1649, #1652
  • DevOps changes:

1.0.0

Full Changelog

This 1.0 release of StellarGraph is the culmination of three years of active research and engineering to deliver an open-source, user-friendly library for machine learning (ML) on graphs and networks.

Jump in to this release, with the new demos and examples:

Major features and improvements

  • Better demonstration notebooks and documentation to make the library more accessible to new and existing users:
  • New algorithms:
    • GCNSupervisedGraphClassification: supervised graph classification model based on Graph Convolutional layers (GCN) #929, demo.
    • DeepGraphCNN (DGCNN): supervised graph classification using a stack of graph convolutional layers followed by SortPooling, and standard convolutional and pooling (such as Conv1D and MaxPool1D) #1212 #1265, demo
    • SortPooling layer: the node pooling layer introduced in Zhang et al #1210
  • DeepGraphInfomax can be used to train almost any model in an unsupervised way, via the corrupt_index_groups parameter to CorruptedGenerator #1243, demo. Additionally, many algorithms provide defaults and so can be used with DeepGraphInfomax without specifying this parameter:
    • any model using FullBatchNodeGenerator, including models supported in StellarGraph 0.11: GCN, GAT, PPNP and APPNP
    • GraphSAGE #1162
    • HinSAGE for heterogeneous graphs with node features #1254
  • UnsupervisedSampler supports a walker parameter to use other random walking algorithms such as BiasedRandomWalk, in addition to the default UniformRandomWalk. #1187
  • The StellarGraph class is now smaller, faster and easier to construct and use:
    • The StellarGraph(..., edge_type_column=...) parameter can be used to construct a heterogeneous graph from a single flat DataFrame, containing a column of the edge types #1284. This avoids the need to build separate DataFrames for each type, and is significantly faster when there are many types. Using edge_type_column gives a 2.6× speedup for loading the stellargraph.datasets.FB15k dataset (with almost 600 thousand edges across 1345 types).
    • StellarGraph's internal cache of node adjacencies is now computed lazily #1291 and takes into account whether the graph is directed or not #1463, and they now use the smallest integer type they can #1289
    • StellarGraph's internal list of source and target nodes are now stored using integer "ilocs" #1267, reducing memory use and making some functionality significantly faster #1444 #1446)
    • Functions like graph.node_features() no longer needs node_type specified if graph has only one node type (this includes classes like HinSAGENodeGenerator, which no longer needs head_node_type if there is only one node type) #1375
  • Overall performance and memory usage improvements since 0.11, in numbers:
    • The FB15k graph has 15 thousand nodes and 483 thousand edges: it is now 7× faster and 4× smaller to construct (without adjacency lists). It is still about 2× smaller when directed or undirected adjacency lists are computed.
    • Directed adjacency matrix construction is up to 2× faster
    • Various samplers and random walkers are faster: HinSAGENodeGenerator is 3× faster (on MovieLens), Attri2VecNodeGenerator is 4× faster (on CiteSeer), weighted BiasedRandomWalk is up to 3× faster, UniformRandomMetapathWalk is up to 7× faster

Breaking changes

  • The stellargraph/stellargraph docker image wasn't being published in an optimal way, so we have stopped updating it for now #1455
  • Edge weights are now validated to be numeric when creating a StellarGraph. Previously edge weights could be any type, but all algorithms that use them would fail with non-numeric types. #1191
  • Full batch layers no longer support an "output indices" tensor to filter the output rows to a selected set of nodes #1204 (this does not affect models like GCN, only the layers within them: APPNPPropagationLayer, ClusterGraphConvolution, GraphConvolution, GraphAttention, GraphAttentionSparse, PPNPPropagationLayer, RelationalGraphConvolution). Migration: post-process the output using tf.gather manually or the new sg.layer.misc.GatherIndices layer.
  • GraphConvolution has been generalised to work with batch size > 1, subsuming the functionality of the now-deprecated ClusterGraphConvolution (and GraphClassificationConvolution) #1205. Migration: replace stellargraph.layer.ClusterGraphConvolution with stellargraph.layer.GraphConvolution.
  • BiasedRandomWalk now takes multi-edges into consideration instead of collapsing them when traversing the graph. It previously required all multi-edges had to same weight and only counted one of them when considering where to walk, but now a multi-edge is equivalent to having an edge whose weight is the sum of the weights of all edges in the multi-edge #1444

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.

  • GraphConvolutionLSTM: time series prediction on spatio-temporal data, combining GCN with a LSTM model to augment the conventional time-series model with information from nearby data points #1085, demo

Bug fixes and other changes

  • Random walk classes like UniformRandomWalk and BiasedRandomWalk can have their hyperparameters set on construction, in addition to in each call to run #1179
  • Node feature sampling was made ~4× faster by ensuring a better data layout, this makes some configurations of GraphSAGE (and HinSAGE) noticeably faster #1225
  • The PROTEINS dataset has been added to stellargraph.datasets, for graph classification #1282
  • The BlogCatalog3 dataset can now be successfully downloaded again #1283
  • Knowledge graph model evaluation via rank_edges_against_all_nodes now defaults to the random strategy for breaking ties, and supports top (previous default) and bottom as alternatives #1223
  • Creating a RelationalFullBatchNodeGenerator is now significantly faster and requires much less memory (18× speedup and 560× smaller for the stellargraph.datasets.AIFB dataset) #1274
  • Creating a FullBatchNodeGenerator or FullBatchLinkGenerator is now significantly faster and requires much less memory (3× speedup and 480× smaller for the stellargraph.datasets.PubMedDiabetes dataset) #1277
  • StellarGraph.info now shows a summary of the edge weights for each edge type #1240
  • The plot_history function accepts a return_figure parameter to return the matplotlib.figure.Figure value, for further manipulation #1309 (Thanks @LarsNeR)
  • Tests now pass against the TensorFlow 2.2.0 release candidates, in preparation for the full 2.2.0 release #1175
  • Some functions no longer fail for some particular cases of empty graphs: StellarGraph.to_adjacency_matrix #1378, StellarGraph.from_networkx #1401
  • CorruptedGenerator on a FullBatchNodeGenerator can be used to train DeepGraphInfomax on a subset of the nodes in a graph, instead of all of them #1415
  • The stellargraph.custom_keras_layers dictionary for use when loading a Keras model now includes all of StellarGraph's layers #1280
  • PaddedGraphGenerator.flow now also accepts a list of StellarGraph objects as input #1458
  • Supervised Graph Classification demo now prints progress update messages during training #1485
  • Explicit contributors file has been removed to avoid inconsistent acknowledgement #1484. Please refer to the GitHub display for contributors instead.
  • Various documentation, demo and error message fixes and improvements: #1141, #1219, #1246, #1260, #1266, #1361, #1362, #1385, #1386, #1363, #1376, #1405 (thanks @thatlittleboy), #1408, #1393, #1403, #1401, #1397, #1396, #1391, #1394, #1434 (thanks @thatlittleboy), #1442, #1438 (thanks @thatlittleboy), #1413, #1450, #1440, #1453, #1447, #1467, #1465 (thanks @thatlittlboy), #1470, #1475, #1480, #1468, #1472, #1474
  • DevOps changes:

1.0.0rc1

Full Changelog

This is the first release candidate for StellarGraph 1.0. The 1.0 release will be the culmination of 2 years of activate development, and this release candidate is the first milestone for that release.

Jump in to this release, with the new demos and examples:

Major features and improvements

  • Better demonstration notebooks and documentation to make the library more accessible to new and existing users:
  • New algorithms:
    • GCNSupervisedGraphClassification: supervised graph classification model based on Graph Convolutional layers (GCN) #929, demo.
  • DeepGraphInfomax can be used to train almost any model in an unsupervised way, via the corrupt_index_groups parameter to CorruptedGenerator #1243, demo. Additionally, many algorithms provide defaults and so can be used with DeepGraphInfomax without specifying this parameter:
    • any model using FullBatchNodeGenerator, including models supported in StellarGraph 0.11: GCN, GAT, PPNP and APPNP
    • GraphSAGE #1162
    • HinSAGE for heterogeneous graphs with node features #1254
  • UnsupervisedSampler supports a walker parameter to use other random walking algorithms such as BiasedRandomWalk, in addition to the default UniformRandomWalk. #1187
  • The StellarGraph class is now smaller, faster and easier to construct:
    • The StellarGraph(..., edge_type_column=...) parameter can be used to construct a heterogeneous graph from a single flat DataFrame, containing a column of the edge types #1284. This avoids the need to build separate DataFrames for each type, and is significantly faster when there are many types. Using edge_type_column gives a 2.6× speedup for loading the stellargraph.datasets.FB15k dataset (with almost 600 thousand edges across 1345 types).
    • StellarGraph's internal cache of node adjacencies now uses the smallest integer type it can #1289. This reduces memory use by 31% on the FB15k dataset, and 36% on a reddit dataset (with 11.6 million edges).

Breaking changes

  • Edge weights are now validated to be numeric when creating a StellarGraph, previously edge weights could be any type, but all algorithms that use them would fail. #1191
  • Full batch layers no longer support an "output indices" tensor to filter the output rows to a selected set of nodes #1204 (this does not affect models like GCN, only the layers within them: APPNPPropagationLayer, ClusterGraphConvolution, GraphConvolution, GraphAttention, GraphAttentionSparse, PPNPPropagationLayer, RelationalGraphConvolution). Migration: post-process the output using tf.gather manually or the new sg.layer.misc.GatherIndices layer.
  • GraphConvolution has been generalised to work with batch size > 1, subsuming the functionality of the now-deprecated ClusterGraphConvolution (and GraphClassificationConvolution) #1205. Migration: replace stellargraph.layer.ClusterGraphConvolution with stellargraph.layer.GraphConvolution.

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.

  • SortPooling layer: the node pooling layer introduced in Zhang et al #1210
  • DeepGraphConvolutionalNeuralNetwork (DGCNN): supervised graph classification using a stack of graph convolutional layers followed by SortPooling, and standard convolutional and pooling (such as Conv1D and MaxPool1D) #1212 #1265
  • GraphConvolutionLSTM: time series prediction on spatio-temporal data, combining GCN with a LSTM model to augment the conventional time-series model with information from nearby data points #1085, demo

Bug fixes and other changes

  • Random walk classes like UniformRandomWalk and BiasedRandomWalk can have their hyperparameters set on construction, in addition to in each call to run #1179
  • Node feature sampling was made ~4× faster by ensuring a better data layout, this makes some configurations of GraphSAGE (and HinSAGE) noticeably faster #1225
  • The PROTEINS dataset has been added to stellargraph.datasets, for graph classification #1282
  • The BlogCatalog3 dataset can now be successfully downloaded again #1283
  • Knowledge graph model evaluation via rank_edges_against_all_nodes now defaults to the random strategy for breaking ties, and supports top (previous default) and bottom as alternatives #1223
  • Creating a RelationalFullBatchNodeGenerator is now significantly faster and requires much less memory (18× speedup and 560× smaller for the stellargraph.datasets.AIFB dataset) #1274
  • StellarGraph.info now shows a summary of the edge weights for each edge type #1240
  • Various documentation, demo and error message fixes and improvements: #1141, #1219, #1246, #1260, #1266
  • DevOps changes:

0.11.1

Full Changelog

This bugfix release contains the same code as 0.11.0, and just fixes the metadata in the Anaconda package so that it can be installed successfully.

Bug fixes and other changes

0.11.0

Full Changelog

Major features and improvements

  • The onboarding/getting-started process has been optimised and improved:
  • New algorithms:
    • Watch Your Step: computes node embeddings by simulating the effect of random walks, rather than doing them. #750.
    • Deep Graph Infomax: performs unsupervised node representation learning #978.
    • Temporal Random Walks (Continuous-Time Dynamic Network Embeddings): random walks that respect the time that each edge occurred (stored as edge weights) #1120.
    • ComplEx: computes multiplicative complex-number embeddings for entities and relationships (edge types) in knowledge graphs, which can be used for link prediction. #901 #1080
    • DistMult: computes multiplicative real-number embeddings for entities and relationships (edge types) in knowledge graphs, which can be used for link prediction. #755 #865 #1136

Breaking changes

  • StellarGraph now requires TensorFlow 2.1 or greater, TensorFlow 2.0 is no longer supported #1008
  • The legacy constructor using NetworkX graphs has been deprecated #1027. Migration: replace StellarGraph(some_networkx_graph) with StellarGraph.from_networkx(some_networkx_graph), and similarly for StellarDiGraph.
  • The build method on model classes (such as GCN) has been renamed to in_out_tensors #1140. Migration: replace model.build() with model.in_out_tensors().
  • The node_model and link_model methods on model classes has been replaced by in_out_tensors #1140 (see that PR for the exact list of types). Migration: replace model.node_model() with model.in_out_tensors() or model.in_out_tensors(multiplicity=1), and model.node_model() with model.in_out_tensors() or model.in_out_tensors(multiplicity=2).
  • Re-exports of calibration and ensembling functionality from the top-level of the stellargraph module were deprecated, in favour of importing from the stellargraph.calibration or stellargraph.ensemble submodules directly #1107. Migration: replace uses of stellargraph.Ensemble with stellargraph.ensemble.Ensemble, and similarly for the other names (see #1107 for all replacements).
  • StellarGraph.to_networkx parameters now use attr to refer to NetworkX attributes, not name or label #973. Migration: for any named parameters in graph.to_networkx(...), change node_type_name=... to node_type_attr=... and similarly edge_type_name to edge_type_attr, edge_weight_label to edge_weight_attr, feature_name to feature_attr.
  • StellarGraph.nodes_of_type is deprecated in favour of the nodes method #1111. Migration: replace some_graph.nodes_of_type(some_type) with some_graph.nodes(node_type=some_type).
  • StellarGraph.info parameters show_attributes and sample were deprecated #1110
  • Some more layers and models had many parameters move from **kwargs to real arguments: Attri2Vec (#1128), ClusterGCN (#1129), GraphAttention & GAT (#1130), GraphSAGE & its aggregators (#1142), HinSAGE & its aggregators (#1143), RelationalGraphConvolution & RGCN (#1148). Invalid (e.g. incorrectly spelled) arguments would have been ignored previously, but now may fail with a TypeError; to fix, remove or correct the arguments.
  • The method="chebyshev" option to FullBatchNodeGenerator, FullBatchLinkGenerator and GCN_Aadj_feats_op has been removed for now, because it needed significant revision to be correctly implemented #1028
  • The fit_generator, evaluate_generator and predict_generator methods on Ensemble and BaggingEnsemble have been renamed to fit, evaluate and predict, to match the deprecation in TensorFlow 2.1 of the tensorflow.keras.Model methods of the same name #1065. Migration: remove the _generator suffix on these methods.
  • The default_model method on Attri2Vec, GraphSAGE and HinSAGE has been deprecated, in favour of in_out_tensors #1145. Migration: replace model.default_model() with model.in_out_tensors().

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.

  • GCNSupervisedGraphClassification: supervised graph classification model based on Graph Convolutional layers (GCN) #929.

Bug fixes and other changes

  • StellarGraph.to_adjacency_matrix is at least 15× faster on undirected graphs #932
  • ClusterNodeGenerator is now noticeably faster, which makes training and predicting with a ClusterGCN model faster #1095. On a random graph with 1000 nodes and 5000 edges and 10 clusters, iterating over an epoch with q=1 (each clusters individually) is 2× faster, and is even faster for larger q. The model in the Cluster-GCN demo notebook using Cora trains 2× faster overall.
  • The node_features=... parameter to StellarGraph.from_networkx now only needs to mention the node types that have features, when passing a dictionary of Pandas DataFrames. Node types that aren't mentioned will automatically have no features (zero-length feature vectors). #1082
  • A subgraph method was added to StellarGraph for computing a node-induced subgraph #958
  • A connected_components method was added to StellarGraph for computing the nodes involved in each connected component in a StellarGraph #958
  • The info method on StellarGraph now shows only 20 node and edge types by default to be more useful for graphs with many types #993. This behaviour can be customized with the truncate=... parameter.
  • The info method on StellarGraph now shows information about the size and type of each node type's feature vectors #979
  • The EdgeSplitter class supports StellarGraph input (and will output StellarGraphs in this case), in addition to NetworkX graphs #1032
  • The Attri2Vec model class stores its weights statefully, so they are shared between all tensors computed by build #1101
  • The GCN model defaults for some parameters now match the GraphConvolution layer's defaults: specifically kernel_initializer (glorot_uniform) and bias_initializer (zeros) #1147
  • The datasets submodule is now accessible as stellargraph.datasets, after just import stellargraph #1113
  • All datasets in stellargraph.datasets now support a load method to create a StellarGraph object (and other information): AIFB (#982), CiteSeer (#989), Cora (#913), MovieLens (#947), PubMedDiabetes (#986). The demo notebooks using these datasets are now cleaner.
  • Some new datasets were added to stellargraph.datasets:
    • MUTAG: a collection of graphs representing chemical compounds #960
    • WN18, WN18RR: knowledge graphs based on the WordNet linguistics data #977
    • FB15k, FB15k_237: knowledge graphs based on the FreeBase knowledge base #977
    • IAEnronEmployees: a small set of employees of Enron, and the many emails between them #1058
  • Warnings now point to the call site of the function causing the warning, not the warnings.warn call inside StellarGraph; this means DeprecationWarnings will be visible in Jupyter notebooks and scripts run with Python 3.7 #1144
  • Some code that triggered warnings from other libraries was fixed or removed #995 #1008, #1051, #1064, #1066
  • Some demo notebooks have been updated or fixed: demos/use-cases/hateful-twitters.ipynb (#1019), rgcn-aifb-node-classification-example.ipynb (#983)
  • The documentation "quick start" guide duplicated a lot of the information in the README, and so has been replaced with the latter #1096
  • API documentation now lists items under their recommended import path, not their definition. For instance, stellargraph.StellarGraph instead of stellargraph.core.StellarGraph (#1127), stellargraph.layer.GCN instead of stellargraph.layer.gcn.GCN (#1150) and stellargraph.datasets.Cora instead of stellargraph.datasets.datasets.Cora (#1157)
  • Some API documentation is now formatted better #1061, #1068, #1070, #1071
  • DevOps changes:

0.10.0

Full Changelog

Major features and improvements

  • The StellarGraph and StellarDiGraph classes are now backed by NumPy and Pandas #752. The StellarGraph(...) and StellarDiGraph(...) constructors now consume Pandas DataFrames representing node features and the edge list. This significantly reduces the memory use and construction time for these StellarGraph objects.

    The following table shows some measurements of the memory use of g = StellarGraph(...), and the time required for that constructor call, for several real-world datasets of different sizes, for both the old form backed by NetworkX code and the new form backed by NumPy and Pandas (both old and new store node features similarly, using 2D NumPy arrays, so the measurements in this table include only graph structure: the edges and nodes themselves):

    dataset nodes edges size old (MiB) size new (MiB) size change time old (s) time new (s) time change
    Cora 2708 5429 4.1 1.3 -69% 0.069 0.034 -50%
    FB15k 14951 592213 148 28 -81% 5.5 1.2 -77%
    Reddit 231443 11606919 6611 493 -93% 154 33 -82%

    The old backend has been removed, and conversion from a NetworkX graph should be performed via the StellarGraph.from_networkx function (the existing form StellarGraph(networkx_graph) is supported in this release but is deprecated, and may be removed in a future release).

  • More detailed information about Heterogeneous GraphSAGE (HinSAGE) has been added to StellarGraph's readthedocs documentation #839.

  • New algorithms:

    • Link prediction with directed GraphSAGE, via DirectedGraphSAGELinkGenerator #871
    • GraphWave: computes structural node embeddings by using wavelet transforms on the graph Laplacian #822

Breaking changes

  • Some layers and models had many parameters move from **kwargs to real arguments: GraphConvolution, GCN. #801 Invalid (e.g. incorrectly spelled) arguments would have been ignored previously, but now may fail with a TypeError; to fix, remove or correct the arguments.
  • The stellargraph.data.load_dataset_BlogCatalog3 function has been replaced by the load method on stellargraph.datasets.BlogCatalog3 #888. Migration: replace load_dataset_BlogCatalog3(location) with BlogCatalog3().load(); code required to find the location or download the dataset can be removed, as load now does this automatically.
  • stellargraph.data.train_test_val_split and stellargraph.data.NodeSplitter have been removed. #887 Migration: this functionality should be replaced with pandas and sklearn (for instance, sklearn.model_selection.train_test_split).
  • Most of the submodules in stellargraph.utils have been moved to top-level modules: stellargraph.calibration, stellargraph.ensemble, stellargraph.losses and stellargraph.interpretability #938. Imports from the old location are now deprecated, and may stop working in future releases. See the linked issue for the full list of changes.

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.

  • Temporal Random Walks: random walks that respect the time that each edge occurred (stored as edge weights) #787. The implementation does not have an example or thorough testing and documentation.
  • Watch Your Step: computes node embeddings by simulating the effect of random walks, rather than doing them. #750. The implementation is not fully tested.
  • ComplEx: computes embeddings for nodes and edge types in knowledge graphs, and use these to perform link prediction #756. The implementation hasn't been validated to match the paper.
  • Neo4j connector: the GraphSAGE algorithm can execute doing neighbourhood sampling in a Neo4j database, so that the edges of a graph do not have to fit entirely into memory #799. The implementation is not automatically tested, and doesn't support functionality like loading node feature vectors from Neo4j.

Bug fixes and other changes

  • StellarGraph now supports TensorFlow 2.1, which includes GPU support by default: #875
  • Demos now focus on Jupyter notebooks, and demo scripts that duplicate notebooks have been removed: #889
  • The following algorithms are now reproducible:
    • Supervised GraphSAGE Node Attribute Inference #844
    • GraphSAGE Link Prediction #925
  • Randomness can be more easily controlled using stellargraph.random.set_seed #806
  • StellarGraph.edges() can return edge weights as a separate NumPy array with include_edge_weights=True #754
  • StellarGraph.to_networkx supports ignoring node features (and thus being a little more efficient) with feature_name=None #841
  • StellarGraph.to_adjacency_matrix now ignores edge weights (that is, defaults every weight to 1) by default, unless weighted=True is specified #857
  • stellargraph.utils.plot_history visualises the model training history as a plot for each metric (such as loss) #902
  • the saliency maps/interpretability code has been refactored to have more sharing as well as to make it cleaner and easier to extend #855
  • DevOps changes:

0.9.0

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:

0.8.4

Fixed bugs:

  • Fix DirectedGraphSAGENodeGenerator always hitting TypeError exception. #695

0.8.3

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

0.8.2

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

0.8.1

Fixed bugs:

  • Reverted erroneous demo notebooks.

0.8.0

Full Changelog

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

0.7.3

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

0.7.2

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

0.7.1

Full Changelog

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

0.7.0

Full Changelog

Implemented enhancements:

  • SGC Implementation #361 (PantelisElinas)
  • Updated to support Python 3.7 #348
  • FullBatchNodeGenerator now supports a simpler interface to apply different adjacency matrix preprocessing 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

0.6.1 (1 Apr 2019)

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

0.6.0 (14 Mar 2019)

Implemented 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

0.5.0 (11 Feb 2019)

Implemented new features and enhancements:

  • Added model calibration #326
  • Added GraphConvolution layer, GCN class for a stack of GraphConvolution layers, and FullBatchNodeGenerator class for feeding data into GCN models #318
  • Added GraphSAGE attention aggregator #317
  • Added GraphSAGE MaxPoolAggregator and MeanPoolAggregator #278
  • Added shuffle option to all flow methods for GraphSAGE and HinSAGE generators #328
  • GraphSAGE and HinSAGE: ensure that a MLP can be created by using zero samples #301
  • Handle isolated nodes in GraphSAGE #294
  • Ensure isolated nodes are handled correctly by GraphSAGENodeMapper and GraphSAGELinkMapper #182
  • EdgeSplitter: introduce a switch for keeping the reduced graph connected #285
  • Node2vec for weighted graphs #241
  • Fix edge types in demos #237
  • Add docstrings to StellarGraphBase class #175
  • Make L2-normalisation of the final embeddings in GraphSAGE and HinSAGE optional #115
  • Check/change the GraphSAGE mapper's behaviour for isolated nodes #100
  • Added GraphSAGE node embedding extraction and visualisation #290

Fixed bugs:

  • Fixed the bug in running demos when no options given #271
  • Fixed the bug in LinkSequence that threw an error when no link targets were given #273

Refactoring:

  • Refactored link inference classes to use edge_embedding_method instead of edge_feature_method #327