[IGNITE-7438] LSQR: Sparse Equations and Least Squares for Lin Regression

• Type: New Feature
• Status: Resolved
• Priority: Major
• Resolution: Done
• Fix Version/s: 2.5
• Component/s: ml
• Assignee: Anton Dmitriev
• Reporter: Yury Babak
• Created: 16/Jan/18 11:21
• Updated: 16/Apr/18 08:26
• Resolved: 09/Feb/18 11:18

Description

This task consists of two parts:

• Implementation of the LSQR iterative solver of systems of linear equations.
• Implementation of the LSQR-based linear regression trainer.

 

Apache Ignite LSQR iterative solver is based on SciPy reference implementation, but it's distributed and can:

• Efficiently work in cases when a data is distributed across a cluster. 
• Utilize all CPU resources by processing different parts of data on different cores.  

These advantages are achieved as result of changing Golub-Kahan-Lanczos Bidiagonalization Procedure procedure which is a core of LSQR algorithm and utilizing features of Partition Based Dataset implementation.

 

LSQR-based linear regression trainer is a trainer that uses the LSQR solver to solve a system of linear equations which represents a linear regression problem.

Issue Links

• Blocked: IGNITE-7437 Partition based dataset implementation
• causes: IGNITE-8268 Add documentation for linear regression (release 2.5)
• links to: GitHub Pull Request #3494
• links to: Reference implementation (SciPy)
• links to: TeamCity Build
• links to: TeamCity Licenses & Javadoc Build