Table2Vec-automated universal representation learning of enterprise data DNA for benchmarkable and explainable...

Area of Science:

• Data Science
• Machine Learning
• Artificial Intelligence

Background:

• Enterprise data is complex, heterogeneous, and siloed, hindering comprehensive understanding and data-driven decision-making.
• Existing enterprise data warehouses and analytics systems are subject, task, and data-specific, creating analytical silos.
• Effective 'whole-of-enterprise' data understanding is a critical challenge in enterprise data science.

Purpose of the Study:

• To introduce Table2Vec, a neural encoder for automated universal representation learning from enterprise DNA.
• To enable automated data characteristics analysis and data quality augmentation.
• To create representative and benchmarkable enterprise data genomes for enterprise-wide and domain-specific learning tasks.

Main Methods:

• Developed Table2Vec, a neural encoder for representation learning on low-quality enterprise data.
• Integrated automated universal representation learning with downstream learning tasks.
• Illustrated Table2Vec on customer data DNA from complex, heterogeneous, multi-relational big tables.

Main Results:

• Table2Vec learns universal customer vector representations that are all-round, representative, and benchmarkable.
• The learned representations support both enterprise-wide and domain-specific learning goals.
• Table2Vec significantly outperforms existing shallow, boosting, and deep learning methods in enterprise analytics.

Conclusions:

• Table2Vec addresses critical limitations of existing representation learning and enterprise analytics systems.
• Automated universal enterprise representation learning provides a foundation for ethical, whole-of-enterprise machine learning.
• The learned enterprise data DNA offers significant research opportunities and applications in data science.