4DBInfer

4DBInfer is an open-source benchmarking tool focused on graph-centric predictive modeling for relational databases (RDBs)

Developed by Amazon’s Shanghai Lablet, it addresses the lack of public RDB benchmarks for training and evaluation in predictive analytics

The tool enables systematic model comparison across four dimensions: datasets, predictive tasks, RDB-to-graph extraction techniques, and graph-based predictive architectures

It supports a range of predictive tasks, such as missing value imputation, tailored to each dataset

Dummy tables and effective subsampling are used to enhance graph connectivity and scalability

Extensive testing shows that graph-based models leveraging full multi-table RDB structures outperform those using single tables or simple joins

The choice of RDB-to-graph extraction strategy significantly impacts model performance, highlighting the need for flexible experimentation