GraphStorm v0.1.1 release

V0.1.1 is a minor release of GraphStorm. In this release, we add SageMaker support for graph construction, GML model training and inference to simplify the graph ML deployment. We also add multiple feature transformation methods in the graph construction pipeline to simplify graph data preparation. We provide a GraphStorm PyPI package for easy installation.

Major features

• Add SageMaker support for GML model training and inference. ( #140 )
• Sagemaker support for single machine graph construction. ( #218 )
• Speedup link prediction evaluation. The evaluation time of link prediction on MAG is reduced from 1.5 hr to 250s. ( #208 )
• Add a tutorial of running GraphStorm on Amazon SageMaker (https://github.com/awslabs/graphstorm/wiki/scale-sagemaker).
• Add a tutorial of co-training GNN with BERT (https://github.com/awslabs/graphstorm/wiki/advanced-language-models).
• Add support of installing GraphStorm library from PyPI (https://github.com/awslabs/graphstorm/wiki/install-env-setup#setup-graphstorm-with-pip-packages).

Enhancements

• Support custom data split in graph construction. ( #41 )
• Support categorial feature transformation in graph construction ( #50 )
• Support Max-Min feature transformation in graph construction ( #299 )
• Support rank gauss feature transformation in graph construction ( #242 )
• Add weighted edge loss in link prediction training ( #63 )
• Support using edge feature in edge classification and edge regression tasks ( #153 )
• Add a profiler to help understand the runtime performance. ( #206 )

API breaking changes

• We change the format of the value of --fanout and --eval-fanout. In v0.1.0 release, GraphStorm expects “relation_type0:fanout0@relation_type1:fanout1,relation_type0:fanout2@relation_type1:fanout3”, if one wants to specify the fanout for different edge type. Now we change it to “srcntype/relation0/dstntype:fanout0@srcntype/relation1/dstntype:fanout1,srcntype/relation0/dstntype:fanout2@srcntype/relation1/dstntype:fanout3”, so that graphs can have edges with same relation type but different source or destination node types.

Contributors

• Da Zheng from AWS
• Xiang Song from AWS
• Jian Zhang from AWS
• Theodore Vasiloudis from AWS
• Prateek M Desai from AWS
• Runjie Ma from AWS
• Israt Nisa from AWS
• Houyu Zhang from Amazon Search
• Haiyang Yu from Texas A&M University
• WeiLin Cong from Penn State University