# The Relational Data Borg is Learning

**Authors:** Dan Olteanu  
**arXiv:** [2008.07864](https://arxiv.org/abs/2008.07864)  
**Date:** 2020-08-18  
**Summary page:** https://www.summarizepaper.com/en/paper/2008.07864v1/

## Key points

- The paper treats machine learning over relational data as a database problem
- Justification for this approach is based on feature extraction queries and computing group-by aggregates
- The approach has been applied to various supervised and unsupervised learning tasks
- Techniques leveraging knowledge about the underlying data can significantly enhance runtime performance of machine learning
- The paper explores theoretical developments related to the algebraic, combinatorial, and statistical structure of relational data processing
- Systems development involving code specialization, low-level computation sharing, and parallelization are explored to reduce complexity and constant factors in learning time
- Extensive collaboration between the author and colleagues from RelationalAI and the FDB research project
- Acknowledgments for contributions from industry partners such as AWS, GCP, Infor Corporation, LogicBlox Inc., Azure, and RelationalAI
- Funding acknowledgments from EPSRC, ERC, and Horizon 2020 program

## Summary

The paper titled "The Relational Data Borg is Learning" by Dan Olteanu presents an approach that treats machine learning over relational data as a database problem. The author justifies this approach based on two observations. Firstly, the input to the learning task often involves a feature extraction query over the relational data. Secondly, the learning task requires computing group-by aggregates. The paper discusses how this approach has been applied to various supervised and unsupervised learning tasks, including ridge linear regression, factorization machines, support vector machines, decision trees, principal component analysis, k-means clustering, and linear algebra over data matrices. The main message of the work is that the runtime performance of machine learning can be significantly enhanced by utilizing techniques that leverage knowledge about the underlying data. These techniques include theoretical developments related to the algebraic, combinatorial, and statistical structure of relational data processing. Additionally, systems development involving code specialization, low-level computation sharing and parallelization are explored to reduce both complexity and constant factors in learning time. The research presented in this paper is a result of extensive collaboration between the author and colleagues from RelationalAI (Mahmoud Abo Khamis, Molham Aref , Hung Ngo , XuanLong Nguyen) and the FDB research project (Ahmet Kara , Milos Nikolic , Maximilian Schleich , Amir Shaikhha Jakub Zavodny Haozhe Zhang). The author acknowledges their contributions as well as thanks other members of the FDB project for providing figures and examples used in the paper. Furthermore ,the author expresses gratitude for support received from industry partners such as Amazon Web Services (AWS), Google Cloud Platform (GCP), Infor Corporation (Infor), LogicBlox Inc., Microsoft Azure (Azure), and RelationalAI . Funding from EPSRC (Engineering and Physical Sciences Research Council) and ERC (European Research Council) is also acknowledged . The project has received additional funding from the European Union's Horizon 2020 research and innovation program under grant agreement No 682588 . Overall , this paper highlights potential for improving machine learning performance by incorporating techniques that exploit characteristics of relational data , provides insights into theoretical & practical advancements in this area .