- Add macOS (Apple Silicon /
arm64, macOS 12+) pip wheels fordgfanddgf-nightly. - Add
dgf-nightlypackage published daily to PyPI. - Add
dgf.learning.train_node_embedding_modelandNodeEmbeddingModel(GraphMAE self-supervised node embeddings). - Add GraphSAGE and GCN architectures (
architecture="graphsage"andarchitecture="gcn"). - Add
model.analyze()for feature importance and structural sensitivity analysis. - Add
dgf.sampling.BigQueryGraphSampleranddgf.sampling.create_bigquery_view_from_spannerfor BigQuery and Spanner Graph sampling. - Add
dgf.transform.canonicalize_graphanddgf.transform.perturb. - Add TensorFlow Hub, TensorFlow Text, and Keras Hub text embedding normalizers.
- Speed up in-memory subgraph sampling, feature gathering, and GATv2 backward pass.
- Speed-up in-process graph neighbor sampler by ~4x.
- Add support for Graph Transformer layers
dgf.jax.layers.HeterogeneousGraphTransformerConfig. Enable them in the high-level API witharchitecture="heterogeneous_graph_transformer",. - Show the model architecture in the
model.describe()widget. - Make the TG-GNN dependency optional (weak dependency).
- Speed-up semi-distributed graph neighbor sampler.