vO.1.2 Release Notes

This release includes some major bug-fixes, several new features and API changes described below.

Blue Graph’s core

PGFrame

Updates to the PGFrame interface include:

  • Added methods:

    • rename_node_properties and rename_edge_properties for changing property names;

    • add_nodes_from_df and add_edges_from_df for additing nodes and edges using dataframes.

  • Added the from_ontology classmethod for importing (e.g. from Webprotege) ontologies as property graphs.

  • Property values that are added to existing properties are now aggregated into sets (and not replaced as it was before).

Backend support

graph-tool

Fix of a major bug occuring during node merging.

Neo4j

Several minor bugfixes of the Neo4j backend were included in this release. In additon, the interfaces of pgframe_to_neo4j has changed:

  • NaN properties are skipped;

  • Node types can be used as Neo4j node labels;

  • Edge types can be used as Neo4j edge relationship types: edges with multiple types result in multiple Neo4j relationships with respective types and their properties replicated (this behaviour is implemented due to the fact that Neo4j relationships can have exactly one relationship type).

Graph preprocessing with BlueGraph

Semantic property encoding

Updates to the encoders:

  • Word2VecModel is renamed to Doc2VecEncoder and is inherited from bluegraph.downstream.Preprocessor;

  • Wrapped scikit-learn’s TfidfVectorizer into TfIdfEncoder inheriting bluegraph.downstream.Preprocessor.

The above-mentioned changes allow using the BlueGraph’s encoders as a part of EmbeddingPipeline).

Downstream tasks with BlueGraph

Similarity API

Similarity processor updates:

  • Smarter handling of elements not existing in the index (when vectors or similar points are requested, None is returned).

  • Segmented Faiss index can be initialized without vectors, the model can be trained on the first call to add.

Embedding pipelines

Embedding pipeline updates:

  • Added basic prediction interface (the run_prediction method);

  • Modified get_similar_points to be able to query similarity for the unknown vectors;

  • Optimized embedding pipeline memory usage: embedding table is not stored explicitly, but is a part of Faiss index.

Services

Embedder

Embedder is a mini-service for retrieving embedding vectors and similar points distributed as a part of BlueGraph. A detailed description of the API can be found here. Two examples can be found in the Embedder API for NCIt term embedding notebook and Embedder API for node embedding.

This release includes the following updates to the service:

  • Embedder app can predict vectors for unseen points, the following formats can be passed on the input

    • raw: raw data as is

    • json_pgframe: a JSON representation of a PGFrame

    • nexus_dataset: endpoint, bucket, resource id and a Nexus token (in the request header), fetches the dataset by resource ID, downloads it and creates a PGFrame (dataset is a JSON representation of a PGFrame).

  • API changes: the POST method for embedding/ and similar-points/ operates on unseen points;

  • Dockerfile fix (smaller image size), dockerignore updates

  • Embedder app can fetch local models from the directory (specified in the configs).