Teaching Old DB Neu(ral) Tricks: Learning Embeddings on Multi-tabular Databases

Garima Gaur, Rajat Singh, Siddhant Arora, Vinayak Gupta, Srikanta Bedathur

Abstract: The lack of robust representation learning techniques tailored for relational data has led to the underwhelming application of ML models to handle database relevant downstream tasks. Recent works that attempt to embed tabular data into a low dimensional latent space have focused solely on Web tables. A relational database is quite different from a Web table corpus and is way more sophisticated. Existing approaches cannot handle the intricacy of relational databases and often fail to learn meaningful embeddings. To this end, we propose an attention based novel learning technique called RelBert, that intelligently computes the context of entities and learns column semantic aware embeddings. We have implemented an end-to-end system, , and have evaluated its performance for two tasks, missing value imputation (MVI) and instance classification. RelBert has reduced the mean rank by ∼ 40% on an average for MVI task compared to the state-of-the-art approach.