Deployment of Machine Learning models to production at the DK Tax Authorities

CopenhagenR - useR Group
CopenhagenR - useR Group
Public group

University of Copenhagen, CSS, room 1.1.18

Øster Farimagsgade 5B · Copenhagen

How to find us

Across from the Botanical Gardens, enter any of the two gates, go up a flight of stairs to the 1st floor and follow the signs to room 1.1.18 (building 1, 1st floor, room 18)

Location image of event venue

Details

Continouos Delivery of Machine Learning models at the Danish Tax Authorities
======================================
by Kristian Hougaard: Machine Learning Engineer at the Danish Tax Authorities and co-owner of www.superstats.dk.

In this workshop we will describe how the Danish Tax Authorities
implemented a continuouos delivery platform for efficient development
and deployment of machine learning models.

CD and DevOps principles are ubiquitous in software development, and are
also becomming more widely used in machine learning development / data analysis.
We show how we used these principles to build a platform, which gives us
deployment to production and test environments, dependency management,
credentials management, access control, logging, life-cycle management,
etc. of machine learning models.

Combining R and Python in Machine Learning Models
======================================
by Lars Kjeldgaard: Data Scientist at the Danish Tax Authorities and author of the following CRAN packages: 'trimmer', 'customsteps', 'modelgrid', 'recorder' and 'dockr'.

The exciting new package 'reticulate' offers a smooth integration between R and Python, that enables you to get the best of both worlds.

In this talk we will go through, how The Danish Tax Authorities have implemented 'reticulate' in their Machine Learning pipeline, which means that
Data Scientists are now free to combine R and Python in (deployed) Machine Learning models.

Training and deploying multiple related Machine Learning models from a single R package
================================================
by Kasper Brink-Jensen: : Data Scientist at the Danish Tax Authorities

In this workshop we will describe how the Danish tax authorities implemented a framework that facilitates training and deployment of
models predicting tax-subject behavior at high resolution. The same framework is used for prediction with a minimum of arguments through
the use of meta-data associated with R-model objects.
This approach ensures no code duplication for training/scoring as well as a uniform data flow in both scenarios.